Stanford Artificial tnteliigence Laboratory Memo AIM-228 Computer Science Department Report No. STAN-CS-74-409 July 1973 The First Ten Years of Artificial Intelligence Research at Stanford Edited by Lest er Earnest ARTIFICIAL INTELLIGENCE PROJECT John McCarthy, Principal Investigator HEURISTIC PROGRAMMING PROJECT Edward Feigenbaum and Joshua Lederberg, Co-principal Investigators Sponsored by ADVANCED RESEARCH PROJECTS AGENCY ARPA Order No. 457 COMPUTER SCIENCE DEPARTMENT Stanford University
The First Ten Years of Artificial Intelligence Research at Stanford
Edited by -Lester Earnest
ARTIFICIAL INTELLIGENCE PROJECTJohn McCarthy, Principal Investigator
HEURISTIC PROGRAMMING PROJECTEdward Feigenbaum and Joshua Lederberg,
CoAprincipal Investigators
ABSTRACT
The first ten years of research in artificial intelligence and related fields at Stanford Universityhave yielded significant results in computer vision and control of manipulators, speechrecognition, heuristic programming, representation theory, mathematical theory of computation,and modeling of organic chemical processes. This report summarizes the accomplishments andprovides bibliographies in each research area.
This research runs supported by the Advanced Research Projects Agency of the Department ofDefense under Contract SD-153. The views and conclusions contained in this document are thoseof the authors and should not be interpreted as necessarily representing the official policies, eitherexpressed or implied, of the Advanced Research Projects Agency or the U. S. Government.
TABLE OF CONTENTS i
Section .
1. INTRODUCTION 1
2. ARTIFICIAL INTELLIGENCEPROJECT
2.1 Robotics 32.1.1 Manipulation 321.2 Vision 5
2.2 Theoretical Studies2.2.1 Mathematical Theory of
.2.4 Natural Language 222.4.1 Speech Recognition 2224.2 Semantics 24
2.5 Programming Languages 2725.1 LISP 282.5.2 FAIL 292.5.3 SAIL 29
2.6 Computer Facilities 302.6.1 Early Development 312.6.2 Hardware 322.6.3 Software 32
2.7 Associated Projects 352.7.1 Higher Mental Functions 352.7.2 Digital Holography 362.7.3 Sound Synthesis 372.7.4 Mars Picture Processing 38
-
Page
2
Sect ion Page
3. HEURISTIC PROGRAMMINGPROJECT 39
3.1 Summary of Aims andAccomplishments 39
3.2 Current Activity 40
3.3 Views Expressed by OthersConcerning DENDRAL 4 1
Appendices
A. -ACCESS TO DOCUMENTATION 47
B. THESES 49
C. FILM REPORTS 53
D. EXTERNAL PUBLICATIONS. 55
E. A. I. MEMO ABSTRACTS 67
1
1. INTRODUCTION
Artificial Intelligence is the experimental andtheoretical study of perceptual andintellectual processes using computers. Itsultimate goal is to understand these processeswell enough to make a computer perceive,understand and act in ways now onlypossible for humans.
In the late 1950s John McCarthy andMarvin Minsky organized the Artificialintelligence Project at M.I.T. That activityand another at Carnegie Tech (nowCarnegie-Mellon University) did much of thepioneering research in artificial intelligence.
In 1962, McCarthy came to StanfordUniversity and initiated another A. I. Projecthere. He obtained financial support for asmall activity (6 persons) from the AdvancedResearch Projects Agency (ARPA) beginningJune 15, 1963.
A Computer Science Department was formedat Stanford in January 1965. By that timethere were 15 people on the Project,including Edward Feigenbaum who had justarrived. Shortly, a decision was made toexpand the activities of the Project,especially in the area of hand-eye research.Additional support was obtained fromARPA and a PDP-6 computer system wasordered. Lester Earnest arrived in late 1965to handle administrative responsibilities ofthe expanding Project.
By the summer of 1966, the Project hadoutgrown available campus space and movedto the D. C. Power Laboratory in thefoothills -above Stanford. The new computersystem was delivered there. Arthur Samueland Jerome Feldman arrived at about thistime and D. Raj. Reddy joined the faculty,having completed his doctorate on speechrecognition as a member of the Project.Several faculty members from other
departments affiliated themselves with theProject, but without direct financial support:Joshua Lederberg (Genetics), JohnChowning and Leland Smith (Music), andAnthony Hearn (Physics).
By early 1968, there were just over 100people on the Project, about half supportedby ARPA. Kenneth Colby and his groupjoined the Project that year, with separatefunding from the National Institute ofMental Health. Other activities subsequentlyreceived some support from the NationalScience Foundation and the NationalAeronautics and Space Administration.Zohar Manna joined the Faculty and theProject, continuing his work in mathematicaltheory of computation.
In June 1973, the Artificial IntelligenceLaboratory (as it is now called) had 128members, with about two-thirds at leastpartially ARPA-supported. Other ComputerScience Faculty members who have receivedsuch support include Robert Floyd, CordellGreen, and Donald Knuth.
The Heuristic Dendral Project (later changedto Heuristic Programming) was formed in1965 under the leadership of EdwardFeigenbaum and Joshua Lederberg. It wasinitially an element of the A. I. Project andconsisted of five or so people for severalyears.
The Heuristic Dendral Project became aseparate organizational entity with its ownARPA budget in January 1970 and alsoobtained some support from the Departmentof Health, Education, and Welfare. It hashad 15 to 20 members in recent months.
The following sections summarizeaccomplishments of the first 10 years (1963-1973). Appendices list external publications,theses, film reports, and abstracts of researchreports produced by our staff.
2
I 2. ARTIFICIAL INTELLIGENCEPROJECT
The work of the Stanford ArtificialIntelligence Project has been basic andapplied research in artificial intelligence andclosely related fields, including computervision, speech recognition, mathematicaltheory of computation, and control of anartificial arm.
Expenditures from ARPA funds over the tenyea.rs beginning June 15, 1963 haveamounted to $9.2 million. About 43% of thiswas for ‘personnel (salaries, wages, benefits),26% for computer and peripheral equipment(purchase, replacement parts, and rental), 8%for other operating expenses, and 23% forindirect costs.
Here is a short list of what we consider tohave been our main accomplishments. Morecomplete discussions and bibliographiesfollow.
Robotics
Development of vision programs for finding,identifying and describing various kinds ofobjects in three dimensional scenes. Thescenes include objects with flat faces and alsocurved objects.
The development of programs formanipulation and assembly of objects fromparts. The latest result is the completeassembly of an automobile water pump.
Speech Recognition
Development of a system for recognition ofcontinuous speech, later transferred toCarnegie-Mellon University and now beingexpanded upon elsewhere.
Heuristic Programming
Our support of Hearn’s work on symboliccomputation led to the development ofREDUCE, now being extended at theUniversity of Utah and widely usedelsewhere.
Work in heuristic programming resulting inLuckham’s resolution theorem prover. Thisis currently about the best theorem prover inexistence, and it puts us in a position to testthe limitations of current ideas aboutheuristics so we can go beyond them.
Representation Theory
Work in the theory of how to representinformation in a computer is fundamentalfor heuristic programming, for languageunderstanding by computers and forprograms that can learn from experience.Stanford has been the leader in this field.
Mathematical Theory of Computation
Our work in mathematical theory ofcomputation is aimed at replacing debuggingby computer checking of proofs thatprograms meet their specifications.McCarthy, Milner, Manna, Floyd, Igarashi,and Luckham have been among the leadersin developing the relevant mathematicaltheory, and the laboratory has developed thefirst actual proof-checking programs that cancheck proofs that actual programs meet theirspecifications. In particular, Robin Milner’sLCF is a practical proof checker for arevised version of Dana Scott’s logic ofcomputable functions.
Research Facilities
We have developed a laboratory with verygood computer and program facilities andspecial instrumentation for the above areas,including a timesharing system with 64online display terminals.
ARTIFICIAL INTELLIGENCE PROJECT 3
We developed a mechanical arm well suitedto manipulation research. It is being copiedand used by other laboratories.
We designed an efficient display keyboardthat has been adopted at several ARPAnetfacilities. We invented a video switch fordisplay systems that is being widely copied.
In the course of developing our facilities, wehave improved LISP, developed an extendedAlgol compiler called SAIL, and created adocument compiler called PUB (used toproduce this report).
Our early Teletype-oriented text editor calledSOS has become an industry standard andour new display editor “E” is much better.
We have written utility programs for thePDP- 10 and made numerous improvementsto time-sharing systems. Many of ourprograms, particularly LISP, SAIL, and SOS,are. used in dozens of other computer centers.
We designed an advanced central processorthat is about 10 times as fast as our PDP-10.In support of this work, we developedinteractive design programs for digital logicthat have since been adopted by otherresearch and industrial organizations.
Trairring
In the 1963-1973 period, 27 members of ourstaff published Ph.D. theses as ArtificialIntelligence Memos and a number of othergraduate students received direct or indirectsupport from the A. I. Project.
The following subsections review principalactivities, _ with references to publishedarticles and reports.
2.1 Robotics
The project has produced several substantialaccomplishments: an automatic manipulationsystem capable of assembling a water pump,a laser ranging apparatus, and programswhich form symbolic descriptions of complexobjects. We have now focused a major efforton robotics in industrial automation.
2.1.1 Manipulatioll
In 1966, we acquired and interfaced aprosthetic arm from Ranch0 Los AmigosHospital. Although it had major mechanicalshortcomings, the software experience wasvaluable. A program was written for pointto point control of arm movements.Computer servoing was used from thebeginning and has proven much moreversatile than conventional analog servoing.A simple system that visually located blocksscattered on a table and sorted them intostacks according to size was operational bythe spring of 1967 [Pingle 19681.
In order to move through crowdedworkspaces, a program was written to avoidobstacles while carrying out arm movements[Pieper 19681. That program was fairlygeneral but rather slow, since it used a localand not very smart search technique to inchits way around obstacles.
A hydraulic arm was designed and builtaccording to stringent criteria of speed andstrength. Kahn developed a minimum-timeservo scheme, which is close to a bang-bangservo [Kahn f 97 f 1. The effect was impressive(even frightening) but hard on theequipment.
The next arm was designed with software inmind [Scheinman 19691. It was completed inDecember 1970, and has proved a goodresearch manipulator; several other groupshave made copies.
4 ARTIFICIAL INTELLIGENCE PROJECT
Arm control software for the new arm wassplit into a small arm-control servo program,which ran in real time mode on our PDP-6computer, and a trajectory planning program[Paul 197 13, written in a higher levellanguage and running on a timesharedPDP- 10.
The arm servo software contained severalnew features: a) feedforward from aNewtonian dynamic model of the arm,including gravity and inertial forces; b)feedback as a critically damped harmonicoscillator including velocity information fromtachometers; c) trajectory modification facilityin the servo program, which allowsconsiderable change from plannedtrajectories to accomod ate contingencyconditions. Compare this control mechanismwith the usual analog servo; the kinematicmodel and computer servo allow changes ofseveral orders of magnitude in the equivalentservo constants, to account for variations ingravity loads and inertial effects.
The arm was designed so that solution forjoint angles from position and orientation issimple and rapid. The techniques apply to awide range of arms; thus our work hassignificance for industrial and other roboticsapplications.
The MOVE-INSTANCE capability [Paul19711 shows a simple form of automatingsome manipulation procedures. The routinechooses a best grasping and departure for aclass of known objects. Models of theseobjects are stored in order to choose allpossible grasping positions -- parallel faces, aface and a parallel edge, etc. The full rangeof possible departure and approach angles isinvestigated and a solution chosen, ifpossible, which allows manipulation in asingle motion, Otherwise, a solution is chosenwhich uses an intermediate position toregrasp the object. Thus, this facilityprovides a method for grasping generalpolyhedra in arbitrary position and
orientation, provided we have a model of theobject.
Arm planning is based on making completemotions, e.g. picking up an object andputting it down. If we pick up an objectarbitrarily, without thought for putting itdown later, we may not be able to put itdown as desired and may need to grasp itagain. Complete trajectories are piecedtogether from various segments, e.g. grasp,departure, mid-segment, approach, release.Trajectories are specified by the user in aninterpretive hand language (HAL) in termsof macros at the level of inserting a screw.Endpoint positions and orientations are oftenspecified by positioning the arm itself in theproper position and recording joint angles, aform of “learning by doing”. The languageprovides facilities for control using touch andforce sensing.
Pump Assembly
Our first major task was assembly of anautomobile water pump. A film which showsthe process in detail is available fordistribution [Pingle and Paul, “AutomatedPump Assembly”]. We describe the task i.nsome detail to show the level ofprogramming.
The pump parts include pump base, cover,gasket and screws. We chose a plausibleindustrial environment with tools in fixedplaces, screws in a feeder, and pump bodyand cover on a pallet. It is located by vision,then moved by the arm to a standardposition, up against some stops. Pins areinserted into screw holes in the pump bodyto guide alignment of gasket and cover. Ifnecessary the hand searches to seat the pins.
The gasket is placed over the guide pins andvisually inspected for correct placement. Thecover is expected to be within about aquarter of an inch of its fixed place on thepallet. After locating the cover by touch, the
ARTIFICIAL INTELLIGENCE PROJECT
hand places the cover over the guide pins.The hand then picks up a hex head powerscrewdriver, picks up a screw from the feederand inserts the screw and, if necessary,searches to seat the screw. A second screw isinserted. The hand then removes the twopins and inserts four more screws, completingthe assembly. Finally, it tests that the pumpimpeller turns freely.
Three forms of feedback are used, visual,tactile, and force. The visual feedback isprovided by strictly special purpose programswhich have no general interest. We plan togeneralize visual feedback capabilities andinclude them in a system like that used toprogram the rest of the task.
The manipulation parts of the assembly wereprogrammed in the hand language, HAL.The actual program follows:
BEGIN PUtlPALIGN ialign pump base at stopsP I N Pl Hl ;put p i n Pl a t hola Hl
‘PIN P2 H2 ;put pin P2 at holo H 2GFISKETTOPSCREW1 lput i n f i r s t 2 scrwsU N P I N Hl Pl HlA jrrmovr p i n Pl f r o m holr HlU N P I N HZ P2 H2Fl ;remove p i n P 2 f r o m holo H 2SCREW2 ;Insort last 4 screwTESTEND
Each of the commands is a macro. The taskwas performed in a general purpose handlanguage with a control program which couldbe readily adapted to other manipulators.Thus the system is more than a specialpurpose demonstration.
-
5
2.1.2 Vision
During the past two years our vision efforthas shifted from scenes of blocks to outdoorscenes and scenes of complex objects. Inboth cases, interpretation has made use ofworld models.
A crucial part of our work with complexobjects is our development of suitablerepresentation of shape. This is an areaclosely connected with industrial applicationsof research, since representation of shape isimportant in programming of robots, design,display, visual feedback, assembly, andsymbolic description for recognition.
Wichman assembled the first robotics visualfeedback system [Wichman 19671. It wasdeficient in several ways: only the outer edgesof the blocks were observed, the hand had tobe removed from view when visual checkingwas done, and the technique was limited inits domain of objects.
Gill then improved these aspects byrecognizing hand marks and internal edgesof blocks [Gill 19721. The system was modeldriven in that specifying the, projectedappearance of a corner in a small windowprogrammed the system for a new task. Butthe system was limited to tasks withpolyhedra, e.g. stacking blocks and insertingblocks into holes.
A class of representations of shape has beenapplied to symbolic description of shapes oftoys and tools [Agin 1972, Nevatia 19731.The representation depends on a part/wholedescription of objects in terms of parts whichmay themselves be composed of parts. Thesuccess of any part/whole representationdepends on the utility of the primitive parts;within the representation, dominantprimitives are “generalized cones”,originating from a formalization called“generalized translational invariance”. Theseprimitives are locally defined by an arbitrary
6 ARTIFICIAL INTELLIGENCE PROJECT
cross section and a ipace curve called theax is, along which the cross sections aretranslated normal to the axis.
We have developed laser depth ranginghardware which has been operative sinceJanuary 1971. The device is very simpleand can be replicated now for less than$1000, assuming that a suitably sensitivecamera tube such as a silicon vidicon or solidstate sensor is available. From thatexperimental data, we have obtainedcomplete descriptions of a doll, a toy horse, aglove, etc., in terms of part/whole descriptionsin the representation just described. Thesame techniques could be used for monocularimages, with considerable benefit, in spite ofthe added difficulty of using only monocularinformation. We are now writing programswhich match these graph descriptions forrecognition.
The work on visual feedback depends ondisplay techniques which have beendeveloped here and elsewhere. Aninteractive program GEOMED [Baumgart1972al was developed to allow description ofobjects by building them up from a set ofgeometric primitives. The usually tedioustask of input of descriptions of complexobjects is much simplified. For our purposes,a symbolic line drawing output is necessary; afast hidden line elimination program withsymbolic line output structure has beenwritten.
Working with outdoor scenes introduces newdifficulties: surfaces alre textured, line findersetc., have been specialized to scenes ofpolyhedra. We have made substantialprogress by incorporating new visualmodules such as color region finders with aworld niodel. The first of these efforts[Ba jcsy 19721 included a color region finderand Fourier descriptors of texture. Textureswere described by directionality, contrast,element size and spacing. Theseinterpretations from the Fourier transform
are useful but not always valid, and adiscussion of the limitations of Fourierdescriptors was included. Depth cues wereobtained from the gradient of texture. Theresults of color and texture region growingwere shown, and a simulation made ofcombining these various modes with a worldmodel. An interesting conclusion was thatthree dimensional interpretations and cueswere the most semantic help in sceneinterpretation.
A second project has dealt with outdoorscenes [Yakimovsky 19731. Yakimovskymade a region-based system whichsequentially merges regions based onsemantics of familiar scenes, using two-dimensional image properties. The systemhas an initial stage which is very much likeother region approaches, merging regionsbased on similarity of color and other crudedescriptors, except that it eliminates weakestboundaries first. The second stage introducesa world mode1 and a means of estimating thebest interpretation of the scene in terms ofthat model. The semantic evaluationprovides much better regions than the samesystem without the semantics. It has beendemonstrated on road scenes, for whichrather good segmentations and labellirigshave been achieved. An application was alsomade~ to heart angiograms.
In human and animal peception, stereo andmotion perception of depth are important.Several of these mechanisms have beenprogrammed. Nevatia [ 19731 evaluatedmotion parallax of a moving observer forscenes of rocks. That program tracked bycorrelation subsequent images in a series ofimages with small angle between images.The technique allows large baselines fortriangulation. Another approach to stereohas been carried out on outdoor scenes[Hannah unpublished].
.
An edge operator [Hueckel 19711 gives agood indication, based on local brightness, of
/ ARTIFICIAL INTELLIGENCE PROJECT
whether or not there is an edge near thecenter of a neighborhood. If there is an edge,its position and location are relatively welldetermined.
Differencing techniques have been used tofind changes in scenes [Quam 19721. Colorregion growing techniques have been used.Accomodation of camera parameters tooptimize image quality for each task hasbeen extensively used.
A focussing module allows automaticfocussing: Color identification modules havebeen written; the latest incorporates theRetinex model of Land to achieve a certaincolor constancy, independent of the variationof illuminance spectrum from scene to scene(sunlight, incandescent, xenon arc,flourescent) or within a scene (reflections fromcolored objects). Calibration modules allowmaintaining a system in a well-calibratedstate to maintain reliability, and increasereliability by self-calibration.
l
Programs have been written to understandscenes of blocks given line drawings. Anearly version recognized only outlines ofisolated objects. A later program, [Falk19721, arrived at segmentations into objectsby techniques which were extensions ofGuzman’s, but using lines rather thanregions. This dealt more effectively withmissing and extraneous edges. The systemused very limited prototypes of objects, usingsize in an important and restrictive way toidentify objects and their spatial relations.Missing edges were hypothesized for a lineverifier program to confirm or reject.
Still another model-based program [Grape19731 used models of a parallelipiped and awedge to perform the segmentation intoobjects corresponding to models. It is able tohandle a variety of missing and extraneousedges.
Bibliography
[Agin 19721 G. Agin, Description andRepreseutatiou of Curved Objects, PhDThesis in Computer Science, Stanford A.I. Memo AIM- 173, October 1972.
[Bajcsy 19721 R. Bajcsy, ComputerIdentification of Visual Texture, PhDThesis in Computer Science, Stanford A.I. Memo AIM- 180, October 1972.
[Baumgart 1972al B. Baumgart, GEOMED -A Geometric Editor, May 1972.Stanford A. I. Lab., SAILON-68, 1972.
[Baumgart 1972bl Baumgart, Bruce G.,Winged Edge PolyhedronRepresentation, Stanford A. I. MemoAIM- 179, October 1972.
[Binford 1973al T. Binford, Sellsor Systemsfor Manipulation, E. Heer (Ed), RemotelyManned Systens, Calif. Inst. Tech.,Pasadena, 1973.
[Binford 1973131 T. 0. Binford and Jay M.Tenenbaum, Computer Vision,Computer (IEEE), May 1973.
[Earnest 19671 L. D. Earnest, Choosing anEye for a Computer, Stanford A. I.Memo AIM-51, April 1967.
[Falk 197 II G. Falk, Scene Analysis Based011 Imperfect Edge Data, Proc. IJCAI,Brit. Comp. Sot., London, Sept. 197 1.
[Falk 19721 G. Falk, Iuterpretatiou ofImperfect Line Data as a Three-Dimensional Scene, J. ArtijcialIntelligence, Vol 3, No. 2, 1972.
[Feldman 19691 J. Feldman, et al, TheStallford Hand-Eye Project, Proc. IJCAI,Washington DC., 1969.
The thrust of all these efforts is the use ofmodels for perception.
[Feldman 19701 J.A.Feldman, Getting a
8
Computer to see Simple Scenes, IEEEStudent Journal, Sept. 1970.
[Feldman 197 la] J. Feldman and R. Sproull,System Support for the Stanford Hand-Eye System, Proc. IJCAI, BritishComputer Sot., London, Sept. 1971.
[Feldman 1971bJ J. Feldman, et al, The Useof Vision and Manipulation to Solvethe Instant Insanity Puzzle, Proc. j&AI,Brit. Comp. Sot., London, Sept. 1971.
[Feldman 19721 J. Feldman, et al, RecentDevelopments in SAIL -- Au Algol-Based Language for ArtificialIntelligence, Proc. FJCC, 1972.
[Gill 19721 Aharon Gill, Visual Feedbackand Related Problems irr Computer-controlled Hand-eye Coordination, PhDThesis in EE, Stanford A. I. Memo AIM-178, October 1972.
[Grape 19731 Gunnar R. Grape Model-Based(Intermediate-Level) Computer Vision,PhD thesis in Comp. Sci., Stanford A. I.Memo AIM-201, May 1973.
[Hueckel 19711 M.H. Hueckel, An OperatorWhich Locates Edges in DigitizedPictures, Stanford A. I. Memo AIM-105,December 1969, and J. ACM, Vol. 18,No. 1, January 1971.
[Kahn 19713 M. Kahn and B. Roth, TheNear-Minimum-time Control of Opea-Loop Articulated Kirlematic Chains,Trans. ASME, Sept 1971.
[Kelly 19701 Michael D. Kelly, VisualIdentification of People by Computer,PhD thesis in Comp. Sci., Stanford A. I.Memo AIM-130, July 1970.
[McCarthy 19701 J. McCarthy, ComputerControl of a Hand and Eye, Proc. 3rdAMJ nion Conference on Automatic Control
[Montanari 19691 U. Montanari, ContinuousSkeletons from Digitized Images, J.ACM, October 1969.
[Montanari 1970al U. Montanari, A Note onMinimal Length PolygonalApproximation to a Digitized Contour,Comm. ACM, January 1970.
[Montanari 1970bl U. Montanari, On LimitProperties in Digitization Schemes, J.ACM, April 1970.
[Montanari 197Ocl U.Montanari, SeparableGraphs, Planar Graphs and W’ebGrammars, Information and Control, May1970.
[Montanari 1970dl U. Montanari,Heuristically Guided Search andChromosome Matching, ArtificialIntelligence, Vol 1, No. 4, December 1970.
[Montanari 19’711 U. Montanari, On theOptimal Detection of Curves in NoisyPictures, Comm. ACM, May 1971.
[Nevatia 19731 R.K.Nevatia andT.O.Binford, Structured Descriptions ofComplex Objects, P YOC. IJCAI, StanfordUniversity, August 1973.
[Paul 19691 R.Paul, G.Falk, J.Feldman, TheComputer Representation of SimplyDescribed Scenes, Proc 2nd IllinoisGraphics Con.., Univ. Illinois, April 1969.
[Paul 197 I] R. Paul, Trajectory Control of aComputer Arm, Proc. IJCAI, Brit. Comp.Sci., London, Sept. 197 1.
[Paul 19721 R. Paul, Modelling, TrajectoryCalculation and Servoing of aComputer Controlled Arm, PhD thesisin Comp. Sci., Stanford A. I. Memo AIM-177, Sept. 1972.
ARTIFICIAL INTELLIGENCE PROJECT
[Pieper 19681 Donald-L. Pieper, TheKinematics of Manipulators urldercomputer Control, Stanford A. 1. MemoAIM-72 October 1968.
. [Pingle 19681 K. Pingle, J. Singer, andW.Wichman, Computer Corrtrol of aMechanical Arm through Visual Input,Proc. IFIP Congress 1568.
[Pingle 19701 K. Pingle, Visual Perceptionby a Computer, in Automaticlnterpfetatlon and Classtfication ofImages, Academic Press, New York, 1970.
[Pingle 19721 K. K. Pingle and J. M.Tenenbaum, An Accomodating EdgeFollower, Proc. IJCAI, Brit. Comp. Sot.,London, 1971.
[Quam 19721 L. H. Quam, et al, ComputerInteractive Picture Processing, StanfordA. I. Memo AIM-166, April 1972.
[Scheinman 19691 V. D. Scheinman, Designof a Computer Manipulator, Stanford A.I. Memo AIM-92, June 1969.
[Schmidt 19711 R. A. Schmidt, A study ofthe Real-Time Coutrol of a Computer-Driver1 Vehicle, PhD. thesis in EE,Stanford A. I. Memo AIM-149, August1971.
[Sobel 19701 Irwin Sobel, Camera Modelsand Machine Perceptiorr, Stanford A. I.Memo AIM-121, May, 1970.
[Tenenbaum 19701 J. M. Tenenbaum,Accommodatiou in Computer Visiorr,Stanford A. I. Memo AIM- 134,September 1970.
[Tenenbaum 19711 J. M. Tenenbaum, et al,A Laboratory for Hand-Eye Research,Proc IF IP Congress 1971.
[Wichman 19671 W. Wichman, Use of
-
9
optical Feedback in the ComputerControl of an Arm, Stanford A. I. MemoAIM-56, August 1967.
[Y akimovsky 19721 Y. Y akimovsky and J. A.Feldman, A Semantics-Based DecisionTheoretic Regiou Analyzer Proc. IJCAI,Stanford U., August 1973.
FILMS
Gary Feldman, Butter-finger, 16mm colorwith sound, 8 min., March 1968.
Gary Feldman and Donald Pieper, Avoid,16mm silent, color, 5 minutes, March1969.
Richard Paul and Karl Pingle, InstantInsanity, 16mm color, silent 6 min.,August 1971.
Suzanne Kandra, Motion and Vision, 16mmcolor, sound, 22 min., November 1972.
Richard Paul and Karl Pingle AutomatedPump Assembly, 16mm color, Silent,‘Imin, April 1973.
/ 10 ARTIFICIAL INTELLIGENCE PROJECT
2.2 Theoretical Studies
Members of our project have pioneered inmathematical theory of computation,representation theory, and grammaticalinference. This work is not a propersubcategory of artificial intelligence in that itdeals with problems that are basic to all ofcomputer science.
In addition to ARPA sponsorship of thiswork, the mathematical theory ofcomputation activities received some supportfrom the National Aeronautics and SpaceAdministration and the grammaticalinference group has received a grant fromthe National Science Foundation.
2.2.1 Mathematical Theory ofComputation
The idea that computer scientists shouldstudy computations themselves rather thenjust the notion of computability (i.e. recursiontheory) was suggested in 1963 by McCarthy[ 1, 21. These early papers suggested thatmathematical methods could be used toprove (or disprove) the following propertiesof programs:1. a program is correct,2. a program terminates,3. two programs are equivalent,4. a translation procedure between two
languages is correct, (i.e. it preserves themeaning of a program),
5. optimized programs are equivalent to theoriginal,
6. one program uses less resources thananother and is therefore more efficient.
These are simply technical descriptions of aprogramer’s day to day problems. Thenotion of correctness of a program is just --“How do we know that a particular programsolves the problem that it was intended to?”The usual way of putting it is: “Does myprogram have bugs in it”. A correctmathematical description of what this means
is a central problem in MTC and is agenuine first step in any attempt tomechanize the debugging of programs. Theequivalence of programs is similar in thatuntil there are clear ways of describing whata program does, saying that they “do” thesame thing is impossible. These technicalproblems are now well enough understood sothat serious attempts to apply the results to“real” programs are beginning.
Attempts to formalize these questions haveproceeded along several lines simultaneously.In [t, 61 McCarthy and Mansfield discussednew languages for expressing these notionswere considered. [4] considered a first orderlogic which contained an “undefined” truth-value. This was one way of explaining whatwas meant by computations which didn’tterminate. [S] used a traditional first orderlogic to describe a subset of ALGOL.
In 131 McCarthy proposed that computersthemselves might be used to check thecorrectness of proofs in formal systems, andwas the first to actually construct a programto carry this out. This suggests that onecould check or possibly look for solutions tothe above problems (in the form of proofs insome formal system). As a result a series ofproof checkers has been built. The first isreported in [71.
In 1966 Floyd [8] published his now wellknown method of assigning assertions to thepaths in a flowchart, in order to findverification conditions the truth of whichguarantee the “correctness” of the originalprogram.
McCarthy, Painter and Kaplan [9, 10, 11, 12,13, 141 used the ideas in [4, 81 to prove:1) the correctness of a compiler for arithmetic
expressions,2) the correctness of several compilers for
algal-like programs,3) the equivalence of some algorithms.
ARTIFICIAL INTELLIGENCE PROJECT 11
.
Kaplan also gave some completeness resultsfor a formal system which talks aboutassignment statements [lo], and discussed theequivalence of programs [13, 141. During thistime another proof checker was written byW. Weiher [15].
c
In a series of articles 2. Manna extended andexpanded Floyd’s original ideas. With A.Pneulli [16, 171 he discussed the relationshipbetween the termination, correctness andequivalence of recursively defined functionsand the satisfiability (or unsatisfiability) ofcertain Jirst order formulas. In [ 171 theywork out an example using the 91 function.In [18] Manna extended his ideas to non-deterministic programs. E. Ashcroft and hedid a similar thing for parallel programs in1211.
Manna (following Floyd) describes the effectsof a program by showing what kinds ofrelations must hold among the values of theprogram variables at different points in theexecution of the program. In particularbetween the input and the output. In 13 11Floyd suggests an interactive system fordesigning correct programs. These ideas aresystematized and expanded by Manna in[34]. He and Ashcroft show how to removeGOT0 statements from programs andreplace them by WHILE statements in [331.
P. Hayes [ 181 again attacked the problem ofa three-valued predicate logic, this time witha machine implementation in mind. Thiscoincided with a paper of Manna andMcCarthy [ 191, which used this logic.
About this time (1969) several importantdevelopments occurred which allowed theabove questions to be reexamined fromdifferent points of view.1. In [22] 2. Manna showed how to
formulate the notion of partial correctnessin second logic.
Hoare shows how properties (including themeaning) of a program can be expressed asrules of inference in his formal system andhow these rules can be used to generate therelations described by Floyd and Manna.This puts their approach in a formal settingsuitable for treatment on a computer. Workon this formal system is at present beingaggressively pursued. Igarashi, London, andLuckham E391 have increased the scope ofthe original rules and have programed asystem called VCG (for verification conditiongenerator) which takes PASCAL programstogether with assertions assigned to loops inthe program and uses the Hoare rules toautomatically generate verification conditions,the proof of which guarantee the correctnessof the original program. These sentences arethen given to a resolution theorem prover[261 which tries to prove them.
2. C. A. R. Hoare [24] published a paperdescribing a new formalism forexpressing the meanings of programs in
-terms of input/output relations.3. S. Igarashi [23] gave an axiomatic
description of an ALGOL-like language.4. D. Scott suggested using the typed lambda
calculus for studying MTC and firstdescribed IN 1970 a mathematical modelof Church’s lambda calculus.
There is also a project started by Suzukiunder the direction of Luckham to developprograms to take account. of particularproperties of arithmetic and arrays whentrying to prove the verification conditions.London also produced an informal proof oftwo Lisp compilers [35].
These together with McCarthy’s axiomaticapproach now represent the most importantdirections in MTC research. They expressdifferent points of view towards themeanings (or semantics) of programs.
Igarashi’s formal system [231 differs fromHoare’s in that the rules of inference actdirectly on the programs themselves ratherthan properties of such programs.
Scott’s work assumes that the most suitable
12
meaning for a program is the function whichit computes and essentially ignores how thatcomputation proceeds. The other approachesare more intentional in that:1) they may not necessarily mention that
function explicitly although it mightappear implicitly.
2) they can (and do) consider notions ofmeaning that are stronger than Scott’s.
For example programs might have to have“similar” computation sequences beforeconsidering them equivalent 1251.
A computer program LCF (for “logic forcomputable functions”) has beenimplemented by Milner [261. This logic usesthe typed lambda calculus to defines thesemantics of programs. Exactly how to dothis was worked out by Weyhrauch andMilner 128, 29, 301. In conjunction Neweyworked on the axiomatization of arithmetic,finite sets, and lists in the LCF environment.This work is still continuing. In additionMilner and Weyhrauch worked with Scott onan axiomatization of the type free lambdacalculus. Much of this work was informallysummarized in [321.
McCarthy attempts to give an axiomatictreatment to a programming language bydescribing its abstract syntax in first orderlogic and stating properties of theprogramming language directly as axioms.This approach has prompted Weyhrauch tobegin the design of a new first order logicproof checker based -on natural deduction.This proof checker is expected to incorporatethe more interesting features of LCF and willdraw heavily on the knowledge gained fromusing LCF to attempt to make the new firstorder proof checker a viable tool for use inproving properties of programs.
This work is all being brought together byprojects that are still to a large extentunfinished. They include1. a new version of LCF including a facility
to search for proofs automatically;
ARTIFICIAL INTELLIGENCE PROJECT
2. the description of the language PASCALin terms of both LCF and in first orderlogic (in the style of McCarthy) in orderto have a realistic comparison betweenthese approaches and that of Floyd,Hoare, et al;
3. a continuation of Newey’s work;4. the discussion of LISP semantics in LCF
and an attempt to prove the correctnessof the London compilers in a formal way(this is also being done by Newey);
5. the design of both special purpose anddomain independent proving proceduresspecifically with program correctness inmind;
6. -the design of languages for describingsuch proof procedures;
‘7. the embedding of these ideas in the newfirst order checker.
In addition to the work described above,Ashcroft, Manna, and Pneuli [36], andChandra and Manna E371 have publishedresults related to program schemas. Manna’sforthcoming book [371 will be the firstgeneral reference in this field.
Some of these references appeared first as A.I. memos and were later published i njournals. In such cases both referencesappear in the bibliography.
Bibliography
[l] McCarthy, John, A Basis for aMathematical Theory of Computation,in Biaffort, P., and Herschberg, D., (eds.),Computer Programming and FormalSystems, North-Holland, Amsterdam,1963.
[2] McCarthy, John, Towards aMathematical Theory of Computation,Proc. IF/P Congress 62, North-Holland,Amsterdam, 1963.
[S] McCarthy, John, CheckingMathematical Proofs by Computer, in
ARTIFICIAL INTELLIGENCE PROJECT 13
PYOC. Symp. on Rkcurslve Function Theor?(]%I), American Mathematical Society,1962.
[4] McCarthy, John, Predicate Calculus with‘UNDEFINED’ as a Truth-value,Stanford A. I. Memo AIM-l, March 1963.
[5] McCarthy, John, A Formal Descriptionof a Subset of Algol, Stanford A. I.Memo AIM-24, September 1964; also inSteele, T., (ed.), Formal LanguageDescription Languages, North Holland,Amsterdam, 1966.
[6] Mansfield, R., A Fortnal Systeln ofComputation, Stanford A. I. MemoAIM-25, September 1964.
[71 McCarthy, John, A Proof-checker forPredicate Calculus, Stanford A. I. MemoAIM-27, March 1965.
[8] Floyd, R. W., Assigning Meauiugs to‘programs, Vol. 19, AmericanMathematical Society, 19-32, 1967.
E91 McCarthy, John, and Painter, J.,Correctness of a Compiler forArithmetic Expressions, Stanford A. I.Memo AIM-40, April 1966; also inMathematical Aspects of ComputerScience, Proc. Symposia in AppliedMathematics, Amer. Math. Sot., NewYork, 1967.
[lo] Painter, J., Setnantic Correctness of aCompiler for an Algol-like Language,Stanford A. I. Memo AIM-44, March1967.
[ 1 I] Kaplan, D., Some Completeness Resultsin the Mathetnatical Theory ofComputation, Stanford A. I. MemoAIM-45, October 1966; also in J. ACM,January 1968.
[12] Kaplan, D., Correctness of a Colnpiler
for Algol-like progratns, Stanford A. I.Memo AIM-48, July 1967.
[131 Kaplan, D., A Forlnal TheoryConcerning the Equivalence ofAlgorithms, Stanford A. I. MemoAIM-59, May 1968.
[I41 Kaplan, D., The Formal TheoreticAnalysis of Strong Equivalence forElelneutal Programs, Stanford A. I.Memo AIM-60, June 1968.
[I51 Weiher, William, The PDP-6 ProofChecker, Stanford A. I. Memo AIM-53,June 1967.
[I63 Manna, Zohar, and Pnueli, Amir, TheValidity Probleln of the 91-f unction,Stanford A. I. Memo AIM-68, August1968.
[I71 Manna, Zohar, and Pneuli, Amir,Fortnalizatiolr of Properties ofRecursively Defined Functions, StanfordA. I. Memo AIM-82, March 1969; also inJ. ACM, Vol. 17, No. 3, July 1970.
[I81 Hayes, Patrick J., A Machine-orientedForlnulatiorl of the ExtendedFunctiorlal Calculus, Stanford A. I.Memo AIM-86, June, 1969.
[I91 Manna, Zohar, and McCarthy, John,Properties of Programs and PartialFunction Logic, Stanford A. I. MemoAIM- 100, October 1969; also in Meltzer,B. and Michie, D. (eds.), MachineIntelligence 5, Edinburgh UniversityPress, Edinburgh, 1970.
[201 Manna, Zohar, The Correctness of Non-detertninistic Programs, Stanford A. I.Memo AIM-95, August 1969; also inArti,ficial Intelligence, Vol. 1, No. 1, 1970.
[21] Ashcroft, Edward, and Manna, Zohar,Forlnalizatiorl of Properties of Parallel
14 ARTIFICIAL INTELLIGENCE PROJECT
Programs, Stanford A. I. MemoAIM- 110, February 1970; also in MachineIntelligence 6, Edinburgh UniversityPress, Edinburgh, 1971.
[22] Manna, Zohar, Second-orderMathematical Theory of Computation,Stanford A. I. Memo AIM-l 11, March1970; also in Proc. ACM Symposium onTheory of Computing, May 1970.
[231 Igarashi, Shigeru, Semantics of Algol-like Statements, Stanford A. I. MemoAIM-129, June 1970.
[25] Milner, Robin, An Algebraic Definitionof Simulation between Programs,Stanford A. I. Memo AIM-142, February,197 1; also in Proc. IJCAI, BritishComputer Society, London, 1971.
[26] Allen, John and Luckham, David, An_ Interactive Theorem-proving Program,
Stanford A. I. Memo AIM-103, October1971.
1271 Milner, Robin, Logic for ComputableFunctions; Description of a MachineImplementation, Stanford A. I. MemoAIM-169, May 1972.
[28] Milner, Robin, Implementation andApplicatious of Scott’s Logic forComputable Functions, Proc. ACM Conj.on Proving Assertions about Programs,ACM Sigplan Notices, January 1972.
[293 Milner, Robin and Weyhrauch, Richard,Proving Compiler Correctness in aMechanised Logic, Machine Intelligence7, Edinburgh University Press,Edinburgh, 1972.
Program Semantics aud Correctness ina Mechanized Logic, Proc. USA-JapanComputer Conference, Tokyo, 1972.
[311 Floyd, Robert W., Toward InteractiveDesign of Correct Programs, StanfordA. I. Memo AIM-l 50, September 1971;also in Proc. IFIP Congress Z!QZ.
E323 Manna, Zohar, Ness, Stephen, andVuillemin, Jean, Inductive Methods forProving Properties of Programs,Stanford A. I. Memo AIM-154,November 197 1; also in ACM SigplanNotices, Vol. 7, No. I, January 1972.
1331 Ashcroft, Edward, and Manna, Zohar,- The Translation of ‘GO-TO’ Programs
to ‘WHILE’ Programs, Stanford A. I.Memo AIM-138, January 1971; also inProc. IFIP Congress Ml.
1341 Manna, Zohar, Mathematical Theory ofPartial Correctness, Stanford A. I. MemoAIM- 139, January 1971; also in J. Camp.and Sys. Sci., June 1971.
[35] London, Ralph L., Correctness of TwoCompilers for a LISP Subset, StanfordA. I. Memo AIM- 15 1, October 197 1.
[36] Ashcroft, Edward, Manna, Zohar, andPneuli, Amir, Decidable Properties ofMonadic Functional Schetnas, StanfordA. I. Memo AIM-148, July 1971.
[371 Chandra, Ashok, and Manna, Zohar,Program Schetnas with Equality,Stanford A. I. Memo AIM-158, December1971.
[381 Manna, Zohar, Introduction toMathematical Theory of Computation,McGraw-Hill, New York, 1974 (toappear).
[30] Weyhrauch, Richard and Milner, Robin,[39] Igarashi, Shigeru, London, Ralph,
Luckham, David, Automatic Program
ARTIFICIAL INTELLIGENCE PROJECT
Verification I: A Logical Basis and itsImpletnestation, Stanford A. I. MemoAIM-200, May 1973.
2.2.2 Representation Theory
When we try to make a computer programthat solves a certain class of problem, ourfirst task is to decide what information isinvolved in stating the problem and isavailable to help in its solution. Next wemust decide how this information is to berepresented in the memory of the computer.Only then can we choose the algorithms formanipulating this information to solve ourproblem. Representation theory deals withwhat information we need and how it isrepresented in the computer. Heuristics isconcerned with the structure of the problemsolving algorithms.
In the past, much work in artificialintelligence has been content with a ratherperfunctory approach to representations. Arepresentation is chosen rather quickly for aclass of problems and then all attention isturned to devising, programming, and testingheuristics. The trouble with this approach isthat the resulting programs lack generalityand are not readily modifiable to attack newclasses of problems.
The first goal of representation theory is todevise a general way of representinginformation in the computer. It should becapable of representing any state of partialinformation necessary to solve it. In i958,McCarthy posed the problem of making aprogram with “common sense” inapproximately these terms and suggestedusing sentences in an appropriate formallanguage to represent what the programknows [ 11. The advantage of representinginformation by sentences is that sentenceshave other sentences as logical consequencesand the program can find consequencesrelevant to the goals at hand. Thus,representation of information by sentencesallows the following.
15
1. A person can instruct the system withoutdetailed knowledge of what sentences arealready in memory. That is, theprocedures for solving a problem usinginformation in sentence form do notrequire that the information be in aparticular order, nor even a particulargrouping of information into sentences.All they require is that what to do is alogical consequence of the collection ofsentences.
2. Similar considerations apply toinformation generated by the programitself.
3. Representing information by sentencesseems to be the only clean way ofseparating that information which iscommon knowledge (and so should bealready in the system) from informationabout a particular problem.
On the other hand, because each sentencehas to carry with it much of its frame ofreference, representation of information bysentences is very voluminous. It seems clearthat other forms of information (e.g. tables)must also be used, but the content of theseother forms should be described by sentences.
In the last ten years, considerable progresshas been made in the use of the sentencerepresentation. In the heuristic direction,theorem proving and problem solvingprograms based on J. Allen Robinson’s.resolution have been developed by bothGreen and Luckham, among others [seeSection 2.3.1 I. The theory of how torepresent facts concerning causality, ability,and knowledge for artificial intelligence hasbeen developed mainly by McCarthy and hisstudents.
The early work in this project revolvedaround McCarthy’s “advice taker” ideas [2-71.McCarthy and Hayes [8l restructured theproblem and connected this work to thesubject of philosophical logic.
16 ARTIFICIAL INTELLIGENCE PROJECT
Later related work includes Becker’s semantic [9] Joseph Becker, The Modeling of Simplememory sys tem C9, 101, Sandewall’s halogic and Inductive Processes in arepresentation of natural language Semantic Memory System, Stanford A.information in predicate calculus t 1 11, and I. Memo AIM-77, January 1969; also inHayes’ study of the frame problem 1121. Proc. IJCAI, Washington D. C., 1969.
Bibliography
[ 11 John McCarthy, Programs withCommon Sewe in Proc. 2555 TeddingtonConf. on Mechantsation of ThoughtProcesses, Vol. 1, pp 77-84, H. M.Stationary Office, London, 1960;reprinted in M. Minsky (ed.), Semanticinformation Processing, MIT Press,Cambridge, Mass., 1968.
121 John McCarthy, Situations, Actions, andCausal Laws, Stanford A. I. MemoAIM-Z, July 1963.
[3] F. Safier, ‘The Mikado’ as an AdviceTaker Problem, Stanford A. I. MemoAIM-3, July 1963.
[41 John McCarthy, Programs withCommon Sense, Stanford A. I. MemoAIM-7, September 1963.
[S] M. Finkelstein and F. Safier,Axiomatization and Impletnentation,Stanford A. I. Memo AIM-15, June 1964.
IS] John McCarthy, Formal Description ofthe Game of Pang-ke, Stanford A. I.Memo AIM- 17, July 1964. .
[71 Barbara Huberman, Advice Taker andGPS, Stanford A. I. Memo AIM-33, June1965.
[8] John McCarthy and Patrick Ha.yes, SomePhilosophical Problems front theStandpoint of Artificial Intelligence,Stanford A. I. Memo AIM-73, November1968; also in D. Michie (ed.), MachineIntelligence 4, American Elsevier, NewYork, 1969.
1101 Joseph Becker, An Information-processing Model of Intermediate-levelCognition, Stanford A. I. MemoAIM-l 19, May 1970.
[ill Erik Sandewall, Representing Natural-language Information in PredicateCalculus, Stanford A. I. Memo AIM-128,
_ July 1970.
[I21 Patrick Hayes, The Frame Problem andRelated Problems in ArtificialIntelligence, Stanford A. I. MemoAIM-153, November 1971.
22.3 Grammatical Inference
Professor Feldman and a small group havebeen investigating the p r o b l e m o fgrammatical inference. That is, given a setof strings which has been chosen in arandom way from a formal language such asa context-free language, to make a“reasonable” inference of the grammar forthe language.
Feldman has studied a very general class ofcomplexity measures and shown that the leastcomplex choice of grammar f rom anenumerable class of grammars (which aregeneral rewriting systems), can be found, andhe gives an algorithm for discovering it [3].This approach is also being applied to theinference of good programs for producingspecified input-output behavior.
Bibliography
[ll Jerome A. Feldman, First Thoughts onGrammatical Inference, Stanford A. I.Memo AIM-55, August 1967.
ARTIFICIAL INTELLIGENCE PROJECT 17
[21 Jerome A. Feldman, J. Gips, J. J.Horning, S. Reder, GrammaticalComplexity and Inference, Stanford A.I. Memo AIM-89, June 1969.
[3] Jerome A. Feldman, Some DecidabilityResults 011 Grammatical Inference andComplexity, Stanford A. 1. MemoAIM-93, August 1969; revised May 1970;also in Information and Control, Vol. 20,No. 3, pp. 244262, April 1972.
141 James Jay Horning, A Study ofGrammatical Inference, Stanford A. I.Memo AIM-98, August 1969.
[5] Alan W. Biermann, J. A. Feldman, Onthe Synthesis of Finite-state Acceptors,Stanford A. I. Memo AIM-l 14, April1970.
[6] Alan W. Biermann, On the Inference ofTuriug Machiues from SampleComputations, Stanford A. I. Memo
. AIM- 152, October 1971; also in ArtijkialIntelligence J., Vol. 3, No. 3, Fall 1972.
[7] Alan W. Biermann, On the Synthesis ofFinite-state Machines from Samples oftheir Behavior, lEEE Trans. Computers,Vol. C-2 1, No. 6, June 1972.
[8] Jerome A. Feldman, A. W. Biermann, ASurvey of Grammatical Inference, Proc.ht. Congress on Pattern Recognition,Honolulu, January 1971; also in S.Watanabe (ed.), Frontiers of Pattern-Recognition, Academic Press, 1972.
, [9] Jerome A. Feldman, Paul Shields, TotalComplexity and Inference of BestPrograms, Stanford A. I. MemoAIM-1 59, April 1972.
[lo] Jerome A. Feldman, AutomaticProgramming, Stanford A. I. MemoAIM- 160, February 1972.
2.3 Heuristic Programming
Heuristic programming techniques are a corediscipline of artificial intelligence. Work ontheorem proving, program generation, boardgames, and symbolic computation makeparticularly heavy use of these techniques.An excellent general reference is the book byNilsson [lJ. Nilsson, of SRI, was supportedin part by our project while writing it.
2.3.1 Theorem Proving
The basis of the theorem-proving effort hasbeen the proof procedure for first-order logicwith equality, originally developed by Allenand Luckham 121. This has been extensivelymodified by J. Allen in recent months; thebasic theorem-proving program has beenspeeded up by a factor of 20, an input-outputlanguage allowing normal mathematicalnotation has been a.dded, and the programwill select search strategies automatically ifthe user wishes (we refer to this as automaticmode). As a result it is possible for theprogram to be used by a person who istotally unfamiliar with the theory of theResolution principle and its associated rulesof inference and refinements. A user’smanual is now available [ 101.
This program has been used to obtain proofsof several different research announcementsin the Notices of the American MathematicalSociety, for example, [7, 8, and 91. Morerecently (July 1972) J. Morales learned touse the program essentially by using it toobtain proofs of the results stated in 191 asan exercise. In the course of doing this hewas able to formulate simple ways of usingthe prover to generate possible new theoremsin the same spirit as [91, and did in factsucceed in extending the results of [91.Furthermore, he was able to send the authorsproofs of their results before they hadactually had time to write them up [R. H.Cowen, private correspondence, August 9,19721. Currently, Morales has been applying
18 ARTIFICIAL INTELLIGENCE PROJECT
the theorem-prover to problems in geometry[l 1, 121 that have been the subject of recentpublications in the proceedings of the PolishNational Academy of Sciences. He has beenable to give elementary proofs of some resultsto clarify inaccuracies and omissions. Thiswork is continuing in correspondence withthe authors. The prover is also being usedas part of the program verification system[see Section 22.1 I.
Resolution Prirrciple, SIAM J. Cornput.,Vol. 1, No. 4, December 1972.
[4] David Luckham and Nils J. Nilsson,Ex tractiilg Information f roinResolutioil Proof Trees, ArtificialIntelligence, Vol. 2, No. 1, pp. 27-54,Spring 1971.
A version of the prover, with userdocumentation ElOJ, has been prepared fordistribution to other research groups. Theaddition of a language in which the user canspecify his intuition about how a proof of agiven statement might possibly be obtained,is in progress. J. Allen has alreadyprogrammed a very preliminary version ofthis “HUNCH” language, and has completedthe systems debugging necessary to get acompiled version of Sussman’s CONNIVERlanguage running here. HUNCH languagemay be implemented in CONNIVER, butdiscussions on this point are not yet complete.Initially, it is expected that HUNCH will beuseful in continuing with more difficultapplications in mathematics.
[5] Cordell Green, The Application ofTheorem Proving to QuestioriAnswering Systems, Ph.D. Thesis in E.E., Stanford A. I. Memo AIM-96, August1969.
161 J. Sussman and T. Winograd, MicroPlanner Manual, Project MAC Memo,MIT.
171 Chinthayamma, Sets of IlldependentAxioms for a Teruary Boolean Algebra,Notices Amer. Math. Sot., 16, p. 654, 1969.
[8] E. L. Marsden, A Note OJI ImplicativeModels, Notices Amer. Math. Sot., No.682-02-7, p. 89, January 197 1.
An alternative approach to theorem provingwas developed in Cordell Green’s Thesis onQA3. His work was done at SRI and hisdissertation was published here [Sl.
[9] Robert H. Cowen, Henry Frisz, AlanGrenadir, Some New AXiOJnatiZatiOilSin Group Theory, Preliminary Report,Notices Amer. Math. Sot., No. 72T-112, p.547, June 1972.
[lOI J. R. Allen, Prelimiuary Users Manualfor an Iuterative Theorem-Prover,Stanford Artificial Intelligence LaboratoryOperating Note SAILON-73, 1973.
121 John Allen and David Luckham, AnInteractive Theorem-proving Program,Madine Intelligence 5, Edin burghUniversity Press, Edinburgh, 1970.
[ll] L. Szcerba and W. Szmielew, On theEuclidean Geometry Without the PaschAxiom, Bull. Acad. Polon. Sciences, Ser.Sci. Math., Astronm, Phys., 18, pp. 659-666, 1970.
131 Richard B. Kieburtz and DavidLuckham, Compatibility audComplexity of Refinements of the
1121 Szcerba, L. W., Independence of Pasch’sAxiom, ibid. 131, pp. 491-498.
-
ARTIFICIAL INTELLIGENCE PROJECT 19
2.3.2 Automatic Program Generatio
This work is an outgrowth of an attempt toextend the applicability of the theorem-prover to problems in artificial intelligence,and makes use of a particular notion of aproblem environment (called a “semanticframe”) which was designed for that originalpurpose. However, the system as it nowstands, is independent of the theorem-prover,and is best thought of as a heuristic problem-solver for a subset of Hoare’s logical system.It has been implemented in LISP by J.Buchanan using the backtrack features ofMicro-Planner.
It accepts as input an environment ofprogramming methods and a problem and, ifsuccessful, gives as output a program forsolving the problem. At the moment, theoutput programs are composed of theprimitive operators of the environment,assignments, conditional branches, whileloops, and non-recursive procedure calls.This system has been used to generate manyprograms for solving various problems inrobot control, everyday advice-taking andplanning, and for computing arithmeticalfunctions. It is an interactive facility andincorporates an Advice language whichallows the user to state preferences affectingthe output program and heuristics forspeeding the problem-solving process, andalso to make assumptions.
The system is intended to permit a user tooversee the automatic construction of aprogram according to the current principlesof structured programming. There is also alibrary facility and the system willincorporate library routines into the programunder construction. The details of itsimplementation will be available in JackBuchanan’s Ph.D. Thesis (in preparation).
2.3.3 Board Games
Computer programs that play games such aschess or checkers are deservedly important asresearch vehicles. A Russian computerscientist has said 111 that chess plays the rolein Artificial Intelligence that the fruit flyplays in genetics. Just as the genetics ofDrosophila are studied not to breed betterflies but to study the laws of heredity, so wewrite chess programs, not because it isimportant that computers play good chess,but because chess provides a rather clearbasis for comparing our ideas aboutreasoning processes with humanperformance. Weaknesses in theperformance of the program tell us abouthuman mental processes that we failed toidentify.
John McCarthy supervised the developmentof a chess program at MIT. He brought ithere when he came and subsequentlyimproved it a bit. In 1966, a match was heldwith a program at what was then CarnegieTech. Neither program played especiallywell, but Carnegie eventually resigned. TheStanford program played several games in1967 with a program at the Institute forTheoretical and Experimental Physics inMoscow. Again, both blundered frequently,but ours made some of the worst moves. Byprior agreement, the games were notcompleted because both programs wereknown to have essentially no end-gamecapability.
Barbara Huberman developed a programthat handled certain important end gamesituations and published her dissertation in1968 [51.
Some early work on the game of Kalahproduced a program that played rather wellr2, 31.
Arthur Samuel continued his long-standingdevelopment of a checkers program when he
20 ARTIFICIAL INTELLIGENCE PROJECT
arrived in 1966 [41. His program isapparently still the best in the world. It doesnot defeat the best human players, but playsfairly well against them. Samuelsubsequently decided to apply the “signaturetable learning” scheme that he had developedfor checkers to speech recognition problems[see Section 2.4.11.
Jonathan Ryder developed what was (andprobably still is) the best Go program to date[6]. It plays better than a beginning human,but it is still quite weak by the standards ofexperienced players.
Since 1971, we have done very little work onboard games.
Bibliography
[ 11 A. Kronrod, The Computer becomesMore Intelligent, Isvestiya, March 15,1967; translated in Soviet Cybernetics:Recent New Item, No. 3, Rand Corp.,
- Santa Monica, Calif., April 1967.
[2] R. Russell, Kalah -0 the Game and theProgram, Stanford A. I. Memo AIM-22,September 1964.
[3] R. Russell, Improvemelds to the KalahProgram, Stanford A. I. Memo AIM-23,September 1964.
[4] Arthur Samuel, Some Studies inMachine Learning using the Game ofCheckers, II -- Recent Progress, SianfordA. I. Memo AIM-52, June 1967; also inIBM Journal, November 1967.
[51 Barbara J Huberman, A Program toPlay Chess End Games, Ph.D.Dissertation in Computer Science,Stanford A. I. Memo AIM-65, August1968.
Game of Go, Ph.D. Dissertation inComputer Science, Stanford A. I. MemoAIM- 155, December 197 1.
23.4 Symbolic Computation
The use of computers to manipulatealgebraic expressions and solve systems ofsymbolic equations potentially offerssubstantial improvements in speed andreduced error rate over pencil-and-papermethods. As a consequence, it becomespossible to tackle problems that are out ofthe range of practical human capabilities.
Beginning in 1963, Enea and Wooldridgeworked on the central problem of algebraicsimplification 11, 21. By 1965, Korsvold haddeveloped a working system 131.
At about this time, Hearn became interestedi n the problem because of potentialapplication to problems in particle physics.He developed a system called REDUCE,which was written in LISP [4, 51. Its initialcapabilities included:a) expansion and ordering of rational
functions of polynomials,b) symbolic differentiation,c)- substitutions in a wide variety of forms,d) reduction of quotients of polynomials by
cancellation of common terms,e) calculation of symbolic determinates.
REDUCE has been used for analysis ofFeynman Diagrams and a number of otherproblems in physics and engineering 16, 7,121. It has been extended in a number ofways [8- 11, 13, 141 and is still underdevelopment by Hearn at the University ofUtah.
George Collins spent a sabbatical year here(1972-3) developing his computational system115, 163.
[6] Jonathan L. Ryder, Heuristic Analysisof Large Trees as Generated in the
ARTIFICIAL INTELLIGENCE PROJECT 21
Bibliography
[ 11 Enea, H. and Wooldridge, D. AlgebraicSimplication, Stanford A. I. MemoAIM-5, August 1963.
[2] Wooldridge, D., An Algebraic SimplifyProgram in LISP, Stanford A. I. MemoAIM- 11, December 1963.
[3] Korsvold, K., An On Line AlgebraicSimplificatiorl Program, Stanford A. I.Memo AIM-37, November 1965.
[4] Hearn, A., Computatiou of AlgebraicProperties of Elementary ParticleReactions Using a Digital Computer,Comm. ACM 9, August 1966.
[S] Hearn, A., REDUCE Users’ Manual,Stanford A. I. Memo AIM-50, February1967.
[6] Brodsky, S. and Sullivan, J., W-Boson. Contribution to the AnomalousMagnetic Moment of the Muon, PhysicsReview, 156, 1644, 1967.
[8] Hearn, A., REDUCE, A User-Orientedhteractive System for AlgebraicSimplification, Proceeding5 for ACMSymposium on Interactive Systems for’Experimental Applied Mathematics,August 1967.
[9] Hearn, A., REDUCE, A User-OrientedInteractive System for AlgebraicSimplification, Stanford A. I. MemoAIM-57, October 1967.
[ill Hearn, A., The Problem ofSubstitution, Stanford A. I. MemoAIM- 170, December 1968.
[12] Hearn, A. and Campbell, J.A., SymbolicAnalysis of Feynman Diagrams byComputer, Stanford A. I. Memo AIM-91,August 1969; also in Journal ofComputational Physics 5, 1970.
[13] Hearn, A., Applications of SymbolManipulation in Theoretical Physics,Comm. ACM, August 197 1.
El41 Hearn, A., REDUCE 2, Stanford A. I.Memo AIM- 133, October 1970.
[15] Collins, George, The Computing Timeof the Euclideau Algorithm, Stanford A.I. Memo AIM-187, January 1973.
[161 Collins, George, and Horowitz, Ellis, TheMinimum Root Separation of aPolynomial, Stanford A. I. MemoAIM- 192, April 1973.
[ 101 Hearn, A., The Problem ofSubstitution, Proceedings of IBMSummer Institute on SymbolicMathematics by Computer, July 1968.
22
2.4 Natural Language
We have worked on two aspects of naturallanguage understanding that are still quitedistinct, but may be expected to interconnecteventually.
2.4.1 Speech Recognition
Efforts to establish a vocal communicationlink with a digital computer have beenunderway at Stanford since 1963. Theseefforts have been primarily concerned withfour areas of research:1) basic research in extracting phonemic and
linguistic information from speechwaveforms,
2) the application of automatic learningprocesses,
3) the use of syntax and semantics to aidspeech recognition, and
4) speech recognition systems have been usedto control other processes.
These ,efforts have been carried out inparallel with varying emphasis at differenttimes.
The fruits of Stanford’s speech researchprogram were first seen in October 1964when Raj Reddy published a reportdescribing his preliminary investigations onthe analysis of speech waveforms [ll. Hissystem worked directly with a digitalrepresentation of the speech waveform to dovowel recognition.
By 1966 Reddy had built a much largersystem which obtained a phonemictranscription and which achievedsegmentation of connected phrases utilizinghypotheses tes t ing [ZJ. This systemrepresented a significant contribution towardsspeech sound segmentation [3l. It operatedon a subset of the speech of a singlecooperative speaker.
In 1967 Reddy and his students had refined
ARTIFICIAL INTELLIGENCE PROJECT
several of his processes and published paperson phoneme grouping for speech recognition[4], pitch period determination of speechsounds 151, and computer recognition ofconnected speech 161.
1968 was an extremely productive year forthe Speech group. Pierre Vicens developedan efficient preprocessing scheme for speechanalysis [7I. Reddy and his studentspublished papers on transcription ofphonemic symbols [8l, phoneme-to-graphemetranslation of English [9], segmentation ofconnected speech [lOI, and consonantalclustering and connected speech recognition[l 11. A genera1 paper by John McCarthy,Lester Earnest, Raj Reddy, and PierreVicens described the voice-con trolledartificial arm developed at Stanford [12].
By 1969 the speech recognition processeswere successfully segmenting and parsingcontinuous utterances from a restrictedsyntax [ 13, 14, 151. A short film entitled“Hear Here” was made to document recentaccomplishments.
In mid 1970, Prof. Reddy left Stanford tojoin the faculty of Carnegie-MellonUniversity and Arthur Samuel became thehead of the Stanford speech research efforts.Dr. Samuel had developed a successfulmachine learning scheme which hadpreviously been applied to the game ofcheckers [ 161, [ 171. He resolved to applythem to speech recognition.
By 1971 the first speech recognition systemutilizing Samuel’s learning scheme wasreported by George White [181. This reportwas primarily concerned with theexamination of the properties of signaturetrees and the heuristics involved in theirapplication to an optima1 minimal set offeatures to achieve recognition. Also at thistime, M. M. Astrahan described hishyperphoneme method 1191, which attemptedto do speech recognition by mathematical
ARTIFICIAL INTELLIGENCE PROJECT 23
classifications instead of the traditionalphonemes or linguistic categories. This wasaccomplished bY nearest-neighborclassification in a hyperspace wherein clusterten ters, o r hyperphonemes, had beenestablished.
In 1972 R. B. Thosar and A. L. Samuelpresented a report concerning somepreliminary experiments in speechrecognition using signature tables [201. Thisapproach represented a general attack onspeech recognition employing learningmechanisms at each stage of classification.
The speech effort in 1973 has been devotedto two areas. First, a mathematically rigorousexamination and improvement of thesignature table learning mechanism has beenaccomplished by R. B. Thosar. Second, asegmentation scheme based on signaturetables is being developed to provide accuratesegmentation together with probabilities orconfidence values for the most likelyphoneme occuring during each segment.This process attempts to extract as muchinformation as possible and to pass thisinformation to higher level processes.
In addition to these activities, a new, highspeed pitch detection scheme has beendeveloped by J. A. Moorer and has beensubmitted for publication 1221.
Bibliography
[ 11 D. Raj Reddy, Experiments on .Automatic Speech Recognition by aDigital Computer, Stanford A. I. MemoAI M -26, October 1964.
f21 D. Raj Reddy, AJI Approach toComputer Speech Recognition byDirect Analysis of the SpeechWaveform, Stanford A. I. Memo AIM-43,September 1966.
Sounds, j. Acoust. Sot. Amer., August1966.
141 D. Raj Reddy, Phoneme Grouping forSpeech Recognition, 1. Acoust. sot.Amer., May, 1967.
[Sl D. Raj Reddy, Pitch PeriodDetermination of Speech Sounds, Comm.ACM, June, 1967.
[7] Pierre Vicens, Preprocessing for SpeechAnalysis, Stanford A.I. Memo AIM-71,October 1968.
181 D. Raj Reddy, Computer Transcriptionof Pholiemic Symbols, J. Acoust. Sot.Amer., August 1968.
[9] D. Raj Reddy, and Ann Robinson,Phoneme-To-Grapheme Translation ofEnglish, lEEE Trans. Audio andElectroacoustics, June 1968.
[IO] D. Raj Reddy, and P. Vicens,Procedures for Segmentatiorl ofColinected Speech, J. Audio Eng. Sot.,October 1968.
[I 11 D. Raj Reddy, Consonantal Clusteringand Connected Speech Recognition,Proc. Sixth International Congress ofAcoustics, Vol. 2, pp. C-57 to C-60,Tokyo, 1968.
1121 John McCarthy, Lester Earnest, D. RajReddy, and Pierre Vicens, A ComputerWith Hands, Eyes, and Ears, Proc.FJCC, 1968.
[ 131 Pierre Vicens, Aspects of SpeechRecognition by Computer, Stanford A.I. Memo AIM-85, April 1969.
[31 D. Raj Reddy, Segmentation of Speech
24 ARTIFICIAL INTELLIGENCE PROJECT
[14] D. Raj Reddy, On the IJse ofEnviroumental, Syntactic aridProbabilistic Constraints in Vision andSpeech, Stanford A. I. Memo AIM-78,January 1969.
FILM
[ 151 D. Raj Reddy and R. B. Neely,Contextual Analysis of Phonemes ofEnglish, Stanford A. I. Memo AIM-79,January 1969.
[ 161 A. L. Samuel, Some Studies in MachineLeatniug Using the Game of Checkers,IBM’Journal 3, 21 l-229, 1959.
[ 171 A. L. Samuel, Some Studies in MachineLearning Using the Game of Checkers,II - Recent Progress, IBM Jour. of ReJ.and Dev., 11, pp. 60 l-6 17, November1967.
[18] George M. White, Machine LearningThrough Signature Trees. Applications
_ to Humall Speech, Stanford A. I. MemoAIM- 136, October 1970.
[ 191 M. M. Astrahan, Speech Analysis byClustering, or the Hyper-photiemeMethod, Stanford A. I. Memo AIM-124,June 1970.
[20] R. B. Thosar and A. L. Samuel, SomePreliminary Experiments in SpeechRecognition Using Signature TableLearning, ARPA Speech UnderstandingResearch Croup Note 43, 1972. .
1213 R. B. Thosar, Estimation ofProbability Densities using SignatureTables for Application to PatternRecognition, ARPA SpeechUnderstanding Research Group Note 81.
[22] J. A. Moorer, The Optimum-CombMethod of Pitch Period Analysis inSpeech, Stanford A. I. Memo AIM-207,July 1973.
23. Raj Reddy, Dave Espar, Art Eisenson,Hear Here, 16mm color with sound, 1969.
2.4.2 Semantics
We have two sub-groups working within thegenera1 area of machine translation andnatural language understanding. They arethe Preference Semantics group (Wilks et al.)and the Conceptual Dependency group(&hank et al.)
Both groups stress the centrality of meaningrepresentation and analysis, and theinferential manipulation of suchrepresentations, for the task of naturallanguage understanding. Again, both assumethe importance of context, and of theexpression of meaning in terms of a formalsystem of meaning primitives, while at thesame time denying the centrality ofsyntactical analysis in the conventionallinguist’s sense. Both have capacity forcombining linguistic knowledge withknowledge of the real world, in order to solveproblems of interpretation during the courseof analysis. The two basic formalisms *wereset ‘out definitively in [l l and 141, and thetheoretical basis for systems of this sort hasbeen argued in [9, 11, 131.
The differences between the sub-groups arelargely concerned with the ordering ofresearch goals: the PS [Preference Semantics]group emphasises an immediate task as acriterion of successfu 1 understanding,interlingual machine translation in thepresent case, while the CD [ConceptualDependency] group emphasises independencefrom any particular task. A particularexample of the difference is provided bymachine translation, where, from a givensemantic representation a PS program alwaysconstructs some particular output string forthat representation, whereas a CD programwould want the freedom to construct one of anumber of possible representations.
ARTIFICIAL INTELLIGENCE PROJECT 25
Another, more philosophically motivateddifference of approach between the groups,concerns the time at which inferences are tobe made from semantic representations. Inthe PS approach, no more inferences aremade than is necessary to solve the problemin hand, and the representation is nevermade any “deeper” than necessary. In theCD approach, a large number of inferencesis made spontaneously.
Lastly, the two approaches both aimessentially at machine translation as thefundamental test of a language understander,but the CD approach emphasizes thetreatment of human dialogue while the PSapproach emphasizes the treatment of smalltexts as the definition of a context.
We believe that these approaches arecomplementary rather than exclusive, andthat the parallel testing of different sub-hypotheses in this way, within an overallsemantics-oriented project, can be stimulatingand productive.
Both sub-groups have had pilot systemsrunning on the computer since 19’70, but inthe last year both have implemented muchstronger running programs. During the lastyear the PS implementation has beenenlarged from a basic mode to an extendedm o d e o f the on-line English-Frenchtranslator. It accepts small paragraphs ofEnglish, outputs the semantic representationit derives, and then the French translation.Since it is semantics, rather than syntax,based, the English input need not even begrammatically correct.
The basic mode of Preference Semantics willcope with the problems presented bysentences like “Give the bananas to themonkeys although they are not ripe, becausethey are very hungry”; that is to say, ofattaching the two “they”s correctly on thebasis of what it knows about bananas andmonkeys (which have different genders inFrench, so the choice is important).
The extended mode is called whenreferential problems in a paragraph requirethe manipulation of stronger partialinformation about the real world, as in “Thesoldiers fired at the women and I saw severalfall”, where the referent of “several” dependson our relatively unreliable knowledge ofhow the world works. This is done byextracting inferential information from thebasic representation with common-senseinference rules, that are themselves expressedwithin the semantic notation.
During the last year, the CD group hasbrought toget her three programs toimplement MARGIE: Meaning Analysis,Response Generation and Inference onEnglish. MARGIE is a prototype system,presently undergoing concurrent developmenthere at Stanford, and at the Istituto per glistudi semantici e cognitivi in Castagnola,Switzerland.
MARGIE’s three component programs(analyzer, memorylinferencer, generator) (a)analyze English utterances in context intographs in the deep conceptual baseprescribed by Conceptual Dependency, (b)generate inferences spontaneously from thegraphs and build relational networks whichdemonstrate an ability to understand whatwas said, and (c) generate responses,constructed in this deep conceptual base,back into surface (syntax) networks, andfinally back into syntactically correct Englishsentences. In addition, the analyzer andgenerator may be run coupled directlytogether, providing a paraphraser whichparaphrases the meaning of an Englishsentence rather than simply its syntactic form.
For example, from the input “John killedMary by choking Mary”, the paraphraserwill return with a number of sentencesexamples of which are “John strangledMary” and “John choked Mary and she diedbecause she could not breathe”.
26
In the “Inference mode” the MARGIE systemtakes input such as “John gave Mary abeating with a stick” and produces a numberof inferred output sentences such as “A sticktouched Mary”, “Mary became hurt * and soon.
1.
2.
3.
4.
5.
6.
Bibliography
Goldman, N., Riesbeck, C., AConceptually Based SentenceParaphraser, Stanford ArtificialIntelligence Memo AIM-196, May 1973.
Goldman, N., Computer Generation ofNatural Language from a DeepConceptual Base, Ph.D. Thesis inComputer Science, Stanford University(forthcoming)
Herskovits, A., The generation of Frenchf rain a semantic representation,Stanford Artificial Intelligence MemoAIM-2 12, October 1973.
Rieger, C., Conceptual Memory: ATheory for Processing the MeaningContent of Natural LanguageUtterances, Ph.D. Thesis in ComputerScience, Stanford University(forthcoming).
Riesbeck, C., Computer Analysis ofnatural Language in Context, Ph.D.Thesis in Computer Science, StanfordUniversity (forthcoming).
Schank, R., Goldman, N., Rieger, C.,Riesbeck, C., Primitive ConceptsUnderlying Verbs of Thought, StanfordArtificial Intelligence Memo AIM- 162,Feb. 1972.
Schank, R., Goldman, N., Rieger, C.,Riesbeck, C., MARGIE: Memory,Analysis, Response Generation andInfereuce on English, Advance Papers ofthe Third International Joint Conference
ARTIFICIAL INTELLIGENCE PROJECT
on Artijicial Intelligence, Stanford U.,August 1973.
8. Schank, R., Rieger, C., Inference and theComputer Understanding of NaturalLanguage, Stanford Artificial IntelligenceMemo AIM-197, May 1973.
9. Schank, R., Wilks, Y., The Goals ofLinguistic Theory Revisited, StanfordArtificial Intelligence Memo AIM-202,May 1973.
10. Schank, R., Finding the Conceptual-Content and Intentioti in an Utterancein Natural Language Conversation,Advance Papers of the SecondInternational Joint Conference OnArti$cial Intelligence, British Comp. Sot.,London, 1971.
11. Schank, R., Conceptual Dependency: aTheory of Natural LanguageUnderstanding, Cognitive Psychology,Vol. 3, No. 4, 1972.
12. Schank, R., The Fourteen PrimitiveActions and Their Inferewes, StanfordArtificial Intelligence Memo AIM- 183,February 1973.
13. Wilks, Y., Deci’dability and NaturalLanguage, Mind, December 197 1.
14, Wilks, Y., Grammer, Meaning and theMachine Analysts of Language,Routledge, London, 1972.
15. Wilks, Y., The Stanford MachineTranslation and Understanding Project,in Rustin (ed.), Natural LanguageProcessing, New York, 1973.
16. Wilks, Y., An Artificial IntelligenceApproach to Machine Translation,Stanford Artificial Intelligence MemoAIM-161, February 1972; to appear inSchank and Colby (eds.), Computer
ARTIFICIAL INTELLIGENCE PROJECT
Models of Thought and Language, W. H.Freeman, San Francisco, 1973.
17. Wilks, Y., UnderstaridiJrg WithoutProofs, Advanced Papers of the Thirdinternational Joint Conference onArtijcial Intelligence, Stanford U., August1973.
r 18. Wilks, Y., and Herskovits, A., AnIntelligellt Analyser alid Generator forNatural Language, Proceedings of theInternational Conference onComputational Linguistics, Pisa, 1973.
19. Wilks, Y., Preference Semantics,Stanford Artificial Intelligence MemoAIM-206, July 1973; also in E. Keenan(ed.), PYOC. 1973 Colloquium on FormalSemantics of Natural Language,Cambridge, U.K., 1974 (to appear).
27
2.5 Programming Languages
There is a strong connection between whatyou can say and the language in which yousay it. Advances in artificial intelligence, aswell as other branches of computer science,are frequently linked to advances inprogramming languages. We have found it ’advantageous to invest a part of our effort inlanguage development and the correspondingcompilers and run-time packages.
We have already discussed the developmentof a “hand language” in support of roboticsresearch [Section 2.1.11. This sectiondiscusses more general programminglanguage developments that have taken placein our laboratory.
Feldman and Cries published an analyticalsurvey of translator (e.g. compiler) writingsystems in 1968 [31. Donald Knuthcontinued publication of a series of texts oncomputer programming that has become thestandard reference in the field [4, 5, 61. Healso published reports on language definitionschemes [7], a study of “real world”programming practices [8, summarized in 91and an interesting historical study [ 101.
Bibliography
ill J. Hext, Programming Languages andTranslation, Stanford A. I. MemoAIM- 19, August 1964.
[2] D. Raj Reddy, Source LanguageOptimization of FOR-loops, Stanford A.I. Memo AIM-20, August 1964.
133 Jerome Feldman and David Cries,Trauslator Writillg Systems, Comm.ACM, February 1968.
I41 Donald E. Knuth, The Art of ComputerProgramming, Vol. 1, FundamentalAlgorithms, Addison-Wesley, Menlo Park,Calif., 1968.
28
[5] Donald E. Knuth,‘TRe Art of ComputerProgramming, Vol. 2, SeminumericalAlgorithms, Addison-Wesley, Menlo Park,Calif., 1969.
[6] Donald E. Knuth, The Art of ComputerProgramming, Vol. 3, Sorting andSearching, Addison-Wesley, Menlo Park,Calif., 1973.
[7] Donald E. Knuth, Examples of FormalSemantics, Stanford A. I. MemoAIM-126, July 1970.
[8] Donald E. Knuth, An Empirical Studyof Fortran in Use, Stanford A. I. MemoAIM- 137, November 1970.
[9] Donald E. Knuth, An Empirical Studyof FortraIl Programs, Softzuare --Practice and Experience, Vol. 1, 105- 133,1971.
[IO] Donald E. Knuth, Ancient Babylonian- Algorthms, Comm. ACM, July 1972.
2.5.1 LISP
LISP is the most widely used language inartificial intelligence research. The overalldesign of this programming system wasdefined in John McCarthy’s 1960 article [Ill.LISP 1.5, developed initially at MIT [12],became available on nearly every majorcomputer and remains so today.
Of course, the various versions of “LISP 1.5” -turned out to be not quite compatible. Evenso, Tony Hearn devised a LISP subset that isf a i r ly po r t ab le 1171. This facilitateddistribution of his REDUCE system forsymbolic computation [see Section 23.41.
There was an early attempt to design anadvanced language called LISP 2 [13, 161.The implementation of a compiler was to bedone by a group at System DevelopmentCorperation. Unfortunately, they elaboratedit to death.
ARTIFICIAL INTELLIGENCE PROJECT
Stanford LISP 1.6 was initially developed in1966 from an MIT PDP-6 system. Our staffessentially rewrote the system several timessince iI81 and have distributed it throughDECUS. It is currently in use at dozens ofPDP-6 and PDP- 10 installations.
A group at U. C. Irvine has substantiallyaugmented Stanford LISP in the direction ofBBN LISP. The result is a very convenientand powerful ‘system [ 191.
One of our groups developed the MLISPcompiler [20, 21, 22, 231, which offers anALGOL-like syntax on top of all thestandard LISP features. This language is inuse by investigators in artificial intelligencein various parts of this country and at leastone place in Europe.
More recently, the same group has beenworking on a powerful new language calledLISP70 [241. It allows pattern-directedcomputation and extensibility, i.e. the usercan add his own rewrite rules for newfunctions, which will automatically be mergedand ordered in the existing rules.
1121 John McCarthy, et al, LISP 1.5Programmer’s Manual, MIT Press, 1962.
1131 John McCarthy, Storage Converltiomin LISP 2, Stanford A. I. Memo AIM-8,September 1963.
[I41 S. R. Russell, Improvements in LISPDebugging, Stanford A. I. MemoAIM- 10, December 1963.
[I51 J. Hext, An Expression Input Routinefor LISP, Stanford A. I. Memo AIM-18,July 1964.
ARTIFICIAL INTELLIGENCE PROJECT
[IS] R. W. Mitchell, LISP 2 SpecificationsProposal, Stanford A. I. Memo AIM-21,August 1964.
[I71 Anthony C. Hearn, Standard LISP,Stanford A. I. Memo AIM-90, May 1969.
1181 Lynn Quam, Whitfield Diffie, StanfordLISP 1.6 Manual, Stanford A. I. Lab.Operating note SAILON-28.7, 1973(originally published in 1966).
[ 191 R. J. Bobrow, R. R. Burton, D. Lewis,UC& LISP Manual, U. C. IrvineInformation and Computer ScienceTechnical Report No. 2 1, October 1972.
[201 David C. Smith, MLISP Users’ Manual,Stanford A. I. Memo AIM-84, January1969.
[211 David C. Smith, MLISP, Stanford A. I.Memo AIM-l 35, October 1970.
[22] David C. Smith, Horace Enea, MLISP2,Stanford A. I. Memo AIM-195, May1973.
1231 David C. Smith, Horace Enea,Backtracking in MLISPP, AdvancePapers, International Joint Conf. onArtijkial Intelligence, Stanford U., August1973.
[24] L. G. Tesler, Horace Enea, David C.Smith, The LISP70 Pattern MatchingSystem, Advance Papers, InternationalJoint Conf. on Artijkial Intelligence,Stanford U., August 1973.
29
2.5.2 FAIL
FAIL is a fast one-pass assembler forPDP-10 and PDP-6 machine language. Itwas developed initially by Phil Petit in 19681251 and the fourth revision of the manualwill be published soon [261.
Compared with MACRO-IO, the standardDEC assembler, FAIL uses substantially morememory, but assembles most programs inabout one-fifth the time. FAIL assembles theentire Stanford A. I. timesharing system (twomillion characters) in less than 4 minutes ofCPU time on a KA-10 processor.
FAIL provides more powerful macros thanMACRO and has ALGOL-like blockstructure, permitting local names to be reusedin various parts of a program withoutconflict.
Nearly all our system and utilityprogramming is done in FAIL. It isdistributed through DECUS and is widelyused elsewhere.
Bibliography
[25] P. M. Petit, FAIL, Stanford A. I. Lab.Operating Note SAILON-26, 1968.
E261 F. H. G. Wright, FAIL, Stanford A. I.Memo, in preparation.
2.5.3 SAIL
Members of our staff began work on anALGOL compiler for the PDP-6 in 1967.Work continued intermittently throughDecember 1968, when Dan Swinehart put upGOGOL III [281.
Feldman subsequently guided thedevelopment of an extensively revisedversion that included his LEAP constructs[29, 301. The resulting system, now calledSAIL, has undergone a number of
30 ARTIFICIAL INTELLIGENCE PROJECT
subsequent revisions- (e.g. [311) and a newmanual has been written [321. SAIL is nowa rather sophisticated programming system,offering coroutining, backtracking, contextswitching, and a number of other features.
2.6 Computer Facilities
The development of SAIL has been heavilyinfluenced by the needs of our roboticsresearch. SAIL is in use in a number ofactivities here and at many other installationson the ARPA network.
In support of our research program, we havedeveloped a very efficient display-orientedtimesharing system. The basic hardware is aPDP-lO/PDP-6 dual processor system with256K words of core memory, a Librascopeswapping disk, an IBM 3330 disk file with 85million word capacity, and 64 displayterminals.
Bibliography
[27J John McCarthy (with 12 others),ALGOL 60, Comm. ACM, May 1960 andJan. 1963; Numerische Mathematik,March 1960.
The system is up essentially 24 hours perday, every day; it is not taken down forpreventive maintenance. It comfortablysupports forty-some jobs with a monitor thatis functionally similar to DEC’s TOPS- 10,though quite different inside.
1281 Dan Swinehart, COGOL III, StanfordA. I. Lab. Operating Note SAILON-48,December 1968.
[29] Jerome Feldman, Paul Rovner, The- Leap Larlguage Data Structure, Proc.
IFIP Congress, 1968.
The high performance of the computerfacility results in part from a number oftechnical innovations, both in hardware andsoftware, and a great deal of hard work. Wehave consistently invested about one-third ofour total funding in development of thecomputer system, and consider that ratio tobe about right.
[30] Jerome Feldman, Paul Rovner, AnALGOL-based Associative Language,Stanford A. I. Memo AIM-66, August1968; also in Comm. ACM, August 1969.
[3 11 Jerome Feldman, J. Low, D. C.Swinehart, R. H. Taylor, RecentDevelopmerlts in SAIL, an ALGOL-based language for ArtificialIntelligence, Proc. FJCC, 1972. -
Equipment acquisition costs have been keptrelatively low by not restricting choices todevices that are “compatible” with ourexisting hardware. Instead, we have selectedon a performance/cost basis and usually doneour own interfacing. The resulting total cost,including engineering and technician time, isoften as low as one-third the price of “turn-key” devices, particularly for memories. Ofcourse, the resulting system has a “mosaic”quality, with elements representing nearlyevery manufacturer. Nevertheless, it worksrather well.
[321 Kurt VanLehn (ed.), SAIL UserManual, Stanford A. I. Memo AIM-204,July 1973.
Our current system has the followingfeatures.
1. It successfully combines real time service(e.g. control of mechanical arms) withgeneral timesharing [9, 13, 141. We werethe first to do this.
c
ARTIFICIAL INTELLIGENCE PROJECT 31
2. Everyone in our laboratory has a displayterminal with full graphics capability inhis office. This was made possible bytheir low average cost (about $2K perterminal). We employ a video switch toshare a limited number of displaygenerators a’mong a larger number ofdisplays [ 191. This idea is being copiedin a number of new display systemselsewhere.
These storys are indexed under thewords they contain, so that a person canask for all the stories that mention, say,“Egypt” and have them displayedimmediately. This service has proven tobe very popular with people on theARPA Network.
3. Our display keyboard design usesmultiple shift and control keys to provideboth touch-typing of a large character setand powerful interactive control ofeditors and other system functions. Thiskeyboard is now in use at MIT,Carnegie-Mellon University, and theUniversity of Utah, and other groups areconsidering it.
9. We have a document compiler calledPUB that handles text justification withmultiple variable-width fonts [201. Ifasked, it will automatically handle section,page, and figure numbering, keep cross-references straight, generate a table ofcontents and a sorted index, and otherwonderful things. Pub is in use at nearlyall sites that have acquired XeroxGraphics Printers.
4. Our teletype text editor, called SOS, waswritten some time ago (1968), but still isunsurpassed in its class I1 13. It wasrecently adopted as a standard by theDECUS PDP-10 users group.
5. Our new display editor, called “E”, is evenbetter. Words can’t describe it, but seer2 1, 223.
2.6.1 Early Develop In en t
The concept of a general purposetimesharing system was first proposed byJohn McCarthy when he was at MIT. Thatproposal led to the systems developed inProject MAC. McCarthy also participated inthe development of an early timesharingsystem at BBN E 11.
6. We have a geometric editor, calledCEOMED, for drawing 3-D objectsinteractively and viewing them from anyperspective [ 181.
7. Another family of drawing editors ‘areused to design and check digital logiccircuits and the corresponding printedcircuit cards 1231. This family ofprograms has been adopted by MIT,CMU, and DEC.
Shortly after arriving at Stanford in 1962,McCarthy undertook the development of adisplay-oriented timesharing system, withsupport from the National ScienceFoundation. That system had 12 displayterminals connected to a PDP-1, with a linkto an IBM 7090 [61. An entertaining filmreport summarizes the capabilities of thesystem [71.
8. We have developed a news informationretrieval service called APE [30]. Ourcomputer has a connection to anAssociated Press newswire and maintainsa file of storys from the last 24 hours.
The PDP-1 timesharing system was used forsome early hand-eye experiments with arather crude television camera interface [3,4]and a donated prosthetic arm. The staff whodeveloped the PDP-1 system became thenucleus of the A. I. Project computer groupin 1966, which gave us a head start on thenew system.
32 ARTIFICIAL INTELLIGENCE PROJECT
2.6.2 Hardware
Over ten years, total direct expense forcomputer and experimental equipment (notincluding salaries for construction ormaintenance) was $2.4 million. Of this,about $1.6 million was for capital equipmentpurchases, $400K for interfacing and spareparts, )C2OOK for IBM disk rental, and t200Kfor outside computer time.
Equipment acquisition highlights are asfollows.Date online Equipment1 966 June PDPI6, 64K core, 8 Dectape
1972 Jan. IBM 3330 replaces 2314April Ampex core (128K)May Video Switch
1973 Jan. Xerox Graphics Printer
Beginning in the summer of 1970, we alsoundertook the design of a new high speedprocessor, known locally as “Foonly”, whichwas to be ten times as fast as our PDP-10.The design was completed in early 1973 andfeatured a 2K word cache memory, amicrocode store that was to be accessible bytimeshared users, and a “Console Computer”,which was to be a minicomputer with displayand keyboard. The Console Computer wasto take -the place of the traditional consolelights and switches and, since it was tocontrol the registers and clock of the bigmachine, could be used to monitor anddebug the latter.
The processor was not fabricated, but someof the ideas in it may be expected to appearin commercial equipment. The digital logicdesign programs mentioned earlier weredeveloped in support of this effort.
2.6.3 Software
Our Arst PDP-6 timesharing monitor was theDEC 1.3 Monitor. Since then, we haverewritten essentially all of the system, someparts more than once. We have mostlyadded features; a number of these havefound their way back into later DECmonitors. Thus, we are semi-compatible withcurrent TOPS- 10 monitors.
An area of major revision in our system iskeyboard and display service. We devised a“line editor” for displays, that facilitates rapidentry and modification of lines of text. Usingthe keyboards, we can also control variousformatting parameters of the display and caneven switch to television, to watch televisedseminars on the Stanford campus. Thisservice is a fringe benefit of using standardTV monitors for displays together with avideo switch.
In support of the display service, we had torestructure the monitor to provide freestorage allocation for display data. This freestorage feature has proven useful for anumber of other system functions.
We developed a display-oriented debuggingpackage called RAID [161. Its purpose issimilar to DDT, but the display permits morepowerful interaction, such as maintainingdisplays of selected locations.
When we received the PDP- 10 in 1968, wedevised a dual processor system that permitsthe PDP-10 to be used for both realtime andgeneral timesharing service and the PDP-6 tobe used for realtime only, such as control ofarms and (recently) raster generation for theXerox Graphics Printer.
ARTIFICIAL INTELLIGENCE PROJECT
We completely rewrote the disk file system inthe Monitor, adding redundancy and read-before-write checking. Disk files are backedup by a tape archiving system that gives
excellent protection against both operatingand programming errors [24!. We havenever had a major file loss attributable tosystem hardware or software failure since ourfile system was installed in 1969.
We have developed a resource reservationand allocation system, called RSL 1291, thatallows users to “purchase” reservations forsubsystems and a service level (a percentageof CPU cycles). Reservation costs vary withthe time of day and are stated in a reusablepseudo-currency, barns, whose abundance iscontrolled administratively. The reservationtransactions, enforcement, and accounting areall performed automatically by programs.This permits persons who need assuredservices to have them without worryingabout how many other people climb on thesystem.
An IMP was installed fairly early in the lifeof the ARPA Network and we wrote a telnet
. program to talk to it. At this point wediscovered that the extra core memoryneeded to support this service caused serioussystem performance degradation. As aconsequence, chose not to support networkservices until after we had a full 256K wordsof memory, in the spring of 19’72 [25, 261.
Overall we think that our system has yieldedexcellent returns on the investment, though itsuffers from chronic overloading. CPU timeis the principal bottleneck.
Bibliography
Note that SAILONs listed below areStanford A. I. Lab. Operating Notes, whichare generally written for laboratory internaluse. The text of most SAILONs is kept indisk files in our [S,DOC] area. WorkingNotes are even less formal than SAILONs
-
33
and usually are kept only in the form of textfiles on the disk. Even so, these files arepublic and can be accessed over the ARPANetwork [see Appendix A].
[l] John McCarthy, S. Boilen, E. Fredkin,J.C.R. Licklider, A Time-SharingDebugging System for a SmallComputer, PYOC. AFIPS Con.. (S JCC),Vol. 23, 1963.
[21 John McCarthy, F. Corbato, M. Daggett,The Linking Segment SubprogramLanguage aud Linking LoaderProgramming Languages, Comm. ACM,July ‘963.
[3] P. Carah, A Television CameraInterface for the PDP-1, Stanford A. I.Memo AIM-34, June 1965.
141 J. Painter, Utilization of a TV Camera011 the PDP-1, Stanford A. I. MemoAIM-36, September 1965.
[51 John McCarthy, Time-sharingComputer Systems, in W. Orr (ed.),Conversational Computers, Wiley, 1966.
[6l John McCarthy, D. Brian, C. Feldman,J. Allen, THOR -- A Display BasedTime-sharing System, PYOC. AFIPSConf. (FJCC), Vol. 30, Thompson,Washington D. C., 1967.
[71 Art Eisenson, Gary Feldman, Ellis D.Kropotechev and Zeus, his MarvelousTime-sharing System, 16mm film, blackand white with sound, 15 minutes, March1967.
[8] Richard Cruen, W. Weiher, RapidProgram Generation, Proc. DECUSSymposium, Fall 1968.
191 J. A. Moorer, Dual Processing for thePDP-6/10, Decuscope, Vol 8, No. 3, 1969.
34
[ 101 William Weiher,.Loader Input Format,SAILON-46, October 1968.
[ 111 William Weiher, S. Savitzky, Son ofStopgap, SAILON-50.3, also in disk fileSOS.LES[S,DOC].
[ 121 William Weiher, R. Gruen, RPG --Rapid Program Generator, SAILON-51,1968.
[ 131 J. A. Moorer, Stanford A-I ProjectMonitor Manual: Chapter I - ConsoleCommands, SAILON-54.2, September1970 (revision in preparation).
[ 141 J. A. Moorer, Stanford A-I ProjectMonitor Manual: Chapter II - UserProgramming, SAILON-55.2, September1970 (revision in preparation).
[ 151 Ted Panofsky, Stanford A-I FacilityManual, SAILON-56, May 1973.
[l61 Phil Petit, RAID, SAILON-58.1,February 1970.
[ 173 Richard P. Helliwell, Copy,SAILON-61.1, May 1971.
[181 Bruce Baumgart, GEOMED -- AGeometric Editor, SAILON-68, May1972.
[ 191 Lester Earnest, Video Switch,SAILON-69, May 1972.
[20] Larry Tesler, PUB, the DocumentCompiler, SAILON-70, September 1972.
[21 J Dan Swinehart, TV -- A Display TextEditor, Working Note in disk fileTVED.DCS[UP,DOC], December 1971.
ARTIFICIAL INTELLIGENCE PROJECT
[23] Richard Helliwell, Stanford DrawingProgram, Working Note in disk fileW[F,RPH].
1241 Ralph Gorin, DART -- Dump andRestore Technique, Working Note indisk file DART.REG[UP,DOC 1,November 1972.
[251 Andy Moorer, Telnet Program,Working Note in disk file INET. JAM[UP,DOC].
1261 Dan Swinehart, FTP -- File TransferProtocol, Working Note in disk file
-FTP.DCS[UP,DOC].
[271 Ralph Gorin, Spooler SystemDocumentation, Working Note in diskfile SPOOL.REG[UP,DOC]. February1971. (
1281 Ralph Gorin, Spelling Checker,Working Note in disk fileSPELL.REG[UP,DOC].
[293 Jim Stein, Automatic Service LevelReservation, Working Note in disk fileRSL. JHS[UP,DOC].
[SO1 Martin Frost, Reading Associated PressNews, SAILON-72, July 1973.
[22] Fred Wright, Differences between ‘E’and ‘TV’, Working Note in disk fileTV2E.F W KJP,DOCb
ARTIFICIAL INTELLIGENCE PROJECT 35
2.7 Associated Prdjects Stanford A. I. Memo AIM- 113, March1970.
Our ARPA-sponsored research program hasbenefited from interaction with several
associated but separately supported projects.In addition to intellectual interchange, therehas been some resource sharing among theprojects. For example, some computerequipment purchases and rentals have beenshared on a quid pro quo basis.
[41 Colby, K. M., Mind and Braiu, Again,Stanford A. I. Memo AIM-l 16, March,1970; also in Mathematical BiosciencesVol. 1 1, 47-52, 1970.
2.7.1 Higher Mental Fuuctious
[53 Colby, K. M., and Smith, D. C.,Computer as Catalyst iii the Treatmentof Nonspeaking Autistic Children,Stanford A. I. Memo AIM-120, April1970.
The Higher Mental Functions Project, underthe leadership of Kenneth Colby, issupported by the National Institutes ofHealth. The overall objectives are to aid inthe solution of certain problems in psychiatryand psychotherapy.
[6] Colby, K. M., Weber, S., and Hilf, F. D.,-Artificial Paranoia, Stanford A. I. MemoAIM-125, July 1970; also in ArtificialIntelligence Journal Vol. 2, No. 1, 1972.
One line of research involves a computermodel of paranoia, called PARRY, whichresponds to natural language inquiries [61.Another involves computer-based treatmentof language difficulties in nonspeakingchildren [91. (For this research, the principalinvestigator received t h e 1 9 7 3 FriedaFromm-Reichmann Award from theAmerican Academy of Psychoanalysis.)
[7] Colby, K. M., Hilf, F. D., Weber, S., andKraemer, H. C., A Resemblance Test forthe Validation of a ComputerSimulation of Paranoid Processes,Stanford A. I. Memo AIM-l 56,November 1971.
Bibliography
[ 11 Colby, K. M., and Smith, D. C.,Dialogues betweeu Humans audArtificial Belief Systems, Stanford A. I.Memo AIM-97, August 1969; also in Proc.International Joint Conference on .Artijkiar! Intelligence, 1969.
[8] Colby, K. M., Hilf, F. D., Weber, S., andKraemer, H. C., Turing-likeIndistinguishability Tests for theValidation of a Computer Simulatiollof Paranoid Processes, ArtijcialIntdligence Journal Vol. 3, No. 1, Fall1972.
. [2] Colby, K. M., Tesler, L., and Enea, H. J.,Experiments with a Search Algorithm011 the Data Base of a Humau BeliefStructure, Stanford A. I. Memo AIM-94,August 1969.
[9] Colby, K. M., The Rationale forComputer-based Treatment ofLanguage Difficulties in NonspeakingAutistic Children, Stanford A. I. MemoAIM- 193, April 1973; also in Journal ofAutism and Childhood Schizophrenia, Vol.3, 254-260, 1973.
131 Colby, K. M., Hilf, F. D., and Hall, W. A.,A Mute Patierit’s Experieiice withMachine-Mediated Interviewing,
[lo] Colby, K. M. and Hilf, F. D.,Multidimensional Analysis inEvaluating a Simulation of ParanoidThought, Stanford A. I. Memo AIM- 194,May, 1973.
ARTIFICIAL INTELLIGENCE PROJECT
[ 1 11 Enea, H. J., and Colby, K. M., IdiolectfcLanguage-Analysis for UuderstaudirlgDoctor-Patient Dialogues, AdvancePapers of the Third International JointConference on Artijicial Intelligence, 2%284, 1973.
[12J Hilf, F. D., Colby, K. M., Smith, D. C.,Wittner, W., and Hall, W. A., Machine-Mediated Interviewing, Journal ofNervous and Mental Disease, Vol. 152,No. 4, 1971.
[13] Hilf, F. D., Non-NonverbalCoumunicatioil arid PsychiatricResearch, Archives of General Psychiatry,Vol. 27, November 1972.
[ 143 Hilf, F. D., Partially AutomatedPsychiatric Iuterviewiug -- A ResearchTool, Journal of Nervous and MentalDisease, Vol. 155, No. 6, December 1972.
I153 Schank, R. C., The Fourteeil Primitive. Actions and Their Ilifereuces, Stanford
A. I. Memo AIM-183, February 19’73.
[16] Schank, R. C., and Rieger, C. J. III,Iufereuce aud Computer Uuderstaudilrgof Natural Language, Stanford A. I.Memo AIM-197, May 1973.
[17] Schank, R. C., and Wilks, Y., The Goalsof Linguistic Theory Revisited,Stanford A. I. Memo AIM-202, May1973.
[ 181 Schank, R. C., The Development ofConceptual Structures in Children,Stanford A. I. Memo AIM-203, May,1973.
[ 191 Scfank, R’. C., Goldman, N., Rieger, C.J. III, and Riesbeck, C., Margie: Memory,Analysis, Respouse Generation andIuference OJI English, Advance Papers ofthe Third international Joint Conferenceon Artificial Intelligence, 255-26 1, August1973.
[20] Schank, R. C., Identification ofCouceptualizatious Underlying NaturalLanguage, in R. Schank and K. Colby(eds.), Computer Models of Thought andLanguage, W. H. Freeman, SanFrancisco, 1973.
[21] Schank, R. C., and Colby, K. M., (eds.),Computer Models of Thought andLanguage, W. H. Freeman, SanFrancisco, 1973.
[22] Smith, D. C., and Enea, H. J., MLISPZ,Stanford A. I. Memo AIM-195, May1973.
[23] Smith, D. C., and Enea, H. J.,Backtracking irr MLISPP, AdvancePapers of the Third International JointConference on Artijicial Intelligence, 677-685, August 1973.
[24] Tesler, L., Enea, H. J., and Colby, K. M.,A Directed Graph RepreSeJitatiOli forComputer Simulation of BeliefSystelns, Mathematical Biosciences, Vol.2, 1968.
[251 Tesler, L. G., Enea, H. J. and Smith, D.C., The LISP70 Patteru MatchingSystem, Advance Papers of the ThirdInternational Joint Conference onArtijkial Intelligence, 67 l-676, August1973.
2.7.2 Digital Holography
In a project lead by Joseph Goodman of theElectrical Engineering Department andsponsored by the Air Force, cer ta intechniques for converting holograms intophotographic images were investigated, withpotential applications to satellitephotography.
One interesting byproduct of this work wasthe creation of the world’s first photographtaken without lens or pinhole [2]. A
-
ARTIFICIAL INTELLIGENCE PROJECT
hologram was formed directly on the surfaceof a vidicon television tube, which wasconnected to our computer through adigitizer. The digitized hologram was thenconverted into an image of the originalobject by computer methods and displayedon a CRT.
Bibliography
[l] Joseph Goodman, Digital ImageForin atioil f ram Electrortically DetectedHolograms, in Proc. SPIE Seminar onDig&al imaging Techniques, Sot. Photo-Optical Instrumentation Engineering,Redondo Beach, California, 1967.
[2] Joseph Goodman, Digital ImageFortnatiou from Electrouically DetectedHolograms, Applied Physics Letters,August 1967.
[3] A. Silvestri and J. Goodman, DigitalReCOJlStrUCtiOll of Holographic Images,
John Chowning and Leland Smith of theStanford Music Department and theirstudents have developed computer techniquesfor generating stereophonic andquadraphonic sounds that can beprogrammed to move in two and threedimensions with respect to the listener. Themet hod controls the distribution . andamplitude of direct and reverberant signalsbetween loudspeakers to provide the angularand distance information and introduces aDoppler shift to enhance velocity information131.
Recently, Chowning made an interestingdiscovery that frequency modulationtechniques provide a simple but.effective wayto synthesize certain kinds of sounds [61.
-
37
Leland Smith has developed a graphicseditor capable of handling musical notation,among other things [7l. A number ofcommercial groups had previously tried andfailed to solve this problem.
Bibliography
[ll James Beauchamp (with H. Von Foerster)(eds.), Music by Computers, John Wiley,New York, 1969.
[21 John M. Chowning, The Simulation ofMoving Sound Sources, Proc. AudioEngineering Sot. Convention, May 1970.
[31 James A. Moorer, Music arid ComputerCotnpositiou, Comm. ACM, January1972.
141 Leland Smith, SCORE -- A Musician’sApproach to Computer Music, J. AudioEng. Sot., Jan./Feb. 1972.
[5] James A. Moorer, The HetrodyueMethod of Analysis of TransientWaveforms, Stanford A. I. MemoAIM-208, June 1973.
161 .John M. Chowning, The Synthesis ofComplex Audio Spectra by means ofFrequency Modulatiou, J. AudioEngineering Society, September 1973.
[71 Leland Smith, Editing and PrirltirlgMusic by Computer, J. Music Theory,Fall 1973.
38 ARTIFICIAL INTELLIGENCE PROJECT
2.7.4 Mars Picture Processirlg
As in so many areas, John McCarthy wasamong the first to examine potentialapplications of artificial intelligence toplanetary exploration 113.
[6l Larry Ward, Computer InteractivePicture Processioll, 16 mm color filmwith sound, 8 min., 1972.
More recently, under the sponsorship of theNational Aeronautics and SpaceAdministration, Lynn Quam did adissertation on picture differencingtechniques [2]. The particular set of pictureshe worked on was satellite photographs ofMars containing various geometric andphotometric distortions as well as severalkinds of noise. He successfully solved theproblem of detecting small changes in theplanet surface in the presence of all theseextraneous factors.
His system was subsequently applied topictures of Mars taken by the Mariner 9spacecraft while the mission was in progress[3, 4, 51. A short film shows the interactivedisplay techniques 161.
Bibliography
[ 11 John McCarthy, Computer Control of aMachine for Exploring Mars, StanfordA. I. Memo AIM-14, January 1964.
[2] Lynn H. Quam, Computer Comparisonof Pictures, Stanford A. I. MemoAIM-144, May 1971.
.[31 Lynn H. Quam, et al, ComputerInteractive Picture Processing, StanfordA. I. Memo AIM-166, April 1972.
[41 Carl Sagan, et al, Variable Features 011Mars: Prelitn inary Mariner 9 TelevisiorlResults, Icarus 17, pp. 346-372, 1972.
[ 53 Lynn H. Quam, et al, Mariner 9 PictureDifferencing at Stanford, Sky andTelescope, August 1973.
39
3, HEURISTIC PROGRAMMINGPROJECT
.
T h e Heuristic Programming Projectoriginated in 1965 under the name HeuristicDENDRAL. Its current title reflects abroadening of scope to include several areasof research related by the common theme ofdeveloping high-performance, intelligentprograms for assisting scientific work.
3.1 Summary of Aims atld‘Accomplishments
Heuristic DENDRAL is one of the fewexamples of a high-performance intelligentsystem, sometimes achieving levels o fscientific problem solving not yet achieved bythe best human experts, often achievinglevels equal to that of good humanperformance. However, the attention givento the Heuristic DENDRAL performanceprogram as a successful application ofartificial intelligence research has tended toobscure the more general concerns of theproject investigators. Our aims andaccomplishments have been:
1. To study and construct detailedinformation processing models of processes ofscientific inference. By scientific inference wemean the inferential process by which amodel is constructed to explain a given set ofempirical data. The Heuristic DENDRALsystem is such a model.
2. To study experifnentally the “operatingcharacteristics” and the effectiveness ofdifferent designs (strategies) for thedevelopment of task-specific knowledge in ascientific _ area. The Planning RuleGenerator, a program which takes the theoryused in hypothesis verification and producesrules for guiding hypothesis generation, is aresult of this concern [23]. The GeneralPlanning Program is another example [281.
3. To develop a method for eliciting froman expert the heuristics of scientific judgmentand choice that he is using in theperformance of a complex inference task.We have designed our problem solvingsystems so that the heuristics may beseparated from the programs which usethem. By restricting program design to thistable-driven form [161 new heuristics can beeasily incorporated. Heuristic DENDRAL,Meta-DENDRAL, and the General PlanningProgram employ this methodology.
4. To solve real problems in an area ofsignificance to modern science, and to do sowith-a level of performance high enough tohave a noticeable impact upon that area ofscience. Chemists will agree we have reachedthat stage. For example, the GeneralPlanning Program has been used to analyzemixtures of estrogenic steroids without theneed for gas-chromatographic separation[32]. In the analysis of data for some classesof compounds Heuristic DENDRAL’sperformance matches or exceeds that of apost-doctoral chemist.
5. To discover the heuristics that form thebasis of expert performance. The significantprob!em is not so much tuning a specialistwith new sets of heuristics for new problemsas learning how to acquire these heuristics.The problem of knowledge acquisition andstructuring by problem solving systems iscrucial, and is perhaps the central problem ofAI research today. In recent years we have-made it the main concern of our project.The work on automatic theory formation isfocussed on the development of a programcalled Meta-DENDRAL for automaticacquisition of a theory of fragmentationprocesses in mass spectrometry E33, 341.
6. To study the representation of knowledge.Much of Computer Science can be viewed asa series of programming innovations bymeans of which we are moving graduallyfrom a position of having to tell a computer
40 HEURISTIC PROGRAMMING PROJECT
precisely how we want a problem to besolved to a position of being able to tell aproblem-solving program what we wish doneby the computer. But, what the user wantsdone always concerns some specific taskenvironment--some piece of the real world.For the problem-solving program to interpretfor itself the what, it must have knowledge ofthe specific task environment and i t sbehavior. In other words, the program needssome kind of theory (formal, informal,heuristic, etc.) of that environment in termsof which to do its problem solving. We haveseen in our work that the form in whichknowledge about the (DENDRAL) world isrepresented is crucial to effective problemsolving and to augmenting the knowledgestructure for improved performance. Wehave found the production rule form ofknowledge representation to be flexible, easilyunderstood, manipulable by a computerprogram, and capable of driving a complexproblem solving system 116, 17, 25, 261.
Survey Articles
The research leading to the implementationof the Heuris t ic DENDRAL and Meta-DENDRAL systems has been documented inover thirty publications; a bibliography isincluded at the end of this section. Inparticular, Cl71 and [251 give fairly concisesummaries of the Heuristic DENDRALresearch up through 1970, and reference 29reports on the status of Meta-DENDRAL asof the middle of 1972.
Most Recent Accomplishments
Since 1970, the most significantaccomplishments of the DENDRAL researcheffort have been the design andimplementation of1) an exhaustive and irredundant generator
of topological graphs, thereby extendingHeuristic DENDRAL to cyclic structures[301,
2) a General Planning Program which
3)
interprets high resolution mass spectraldata from complex, biologicallyinteresting molecules having a knownskeletal substructure [28, 321,the initial parts of a theory formationprogram which has already been used toinfer a theory of mass spectrometry for aparticular class of molecules [33J.
3.2 Currelit Activity
At the present time the Project members areworking in the following task areas:
1) Heuristic DENDRAL - Extensions to theHeuristic DENDRAL program are aimedat increasing its utility to practicingchemists by extending its domain ofapplicability to a wider class of molecules.This work is now funded by theBiotechnology Resources Branch of theNational Institutes of Health (Grant No.RR-6 12-O 1).
2) Meta-DENDRAL - Within the domain ofmass spectrometry, a theory formationprogram is under development. The goalis to assist in the inference of the theoryof fragmentation of molecules in a massspectrometer,
3) Intelligent Control of Scientific Instruments- The aim of this research is to developprograms for intelligent control of a data-collecting instrument. T h e p r o g r a mmakes control decisions by reasoning interms of a theory of the instrumental andphysical process involved. This is aproblem of importance when time, bandwidth, or other constraints limit theamount of data which can be gatheredfor analysis. Candidate instruments aremass spectrometers and nuclear magneticresonance spectrometers.
4) Application of AI to the task of computerprogramming (Automatic Programming) -
HEURISTIC PROGRAMMING PROJECT 41
One of the aspects of programming beingworked upon is debugging. In aDENDRAL-like fashion, evidence ofp r0gra.m malfunction (bugs) is“explained” in terms of a model ofcommonly-observed program pathologies.In another effort, the synthesis of portionsof systems programs is the subject of aPh.D. thesis project now beingformulated.
5) Application of AI to a new complex taskdomain of scientific interest, viz. proteinstructure determination from x-raycrystallographic data - Work has recentlybegun to develop heuristic programs thatwill formulate $-space models of thestructure of proteins. The expertise ofprotein chemists in “fitting” complexstructures to poorly resolved and/orincomplete data will be extracted fromexperts and used by the program.
33 Views Expressed by OthersConcerning DENDRAL
T h e DENDRAL publications haveengendered considerable discussion andcomment among computer scientists andchemists. Professor H. Gelernter (SUNY,Stony Brook), at an SJCC 1970 panel of theuse of computers in science gave an extendeddiscussion of the program, in which he saidthat it was the first important scientificapplication of a.rtificial intelligence. Dr. W.H. Ware (RAND Corporation) in a recentarticle entitled “The Computer in YourFuture” in the collection Science andTechnology in the World of the Future (A. B.Bronwell, ed., Wiley Press, 1970) said:
“Thus, much of engineering will beroutinized at a high level ofsophistication, but what aboutscience? An indication of what iscoming a t a higher level ofintellectual performance is a. computerprogram called Heuristic DENDRAL,which does a task that a physical
chemist or biologist concerned withorganic chemistry does repeatedly.”
Professor J. Weizenbaum of MIT, in anundergraduate computer science curriculumproposal for MIT entitled “A First Draft of aProposal for a New Introductory Subject inComputer Science (September 19 70)“,included Heuristic DENDRAL in his “group4” series (three lectures) entitled GreatPrograms; and he said recently (personalcommunication), commenting on recent workand plans,
“I see the work you are nowbeginning as a step in the direction ofcomposing explicit models of justsiich programs (that build expertisein an area). The implications of asuccess in such an effort arestaggering. I therefore believe youreffort to be w o r t h y o f veryconsiderable investment of humanand financial resources.”
In his paper presented at the SixthInternational Machine IntelligenceWorkshop, Professor Saul Amarel (RutgersUniversity) used Heuristic DENDRAL toillustrate a point about programs which usetheoretical knowledge. He wrote:
“The DENDRAL system provides anexcellent vehicle for the study of usesof relevant theoretical knowledge inthe context for forma.tion problems,”from “Representations and Modelingin Problems of Program Formation”,Saul Amarel, in Machine intelligence6 (B. Meltzer and D. Michie, eds.)
_ E d i n b u r g h U n i v e r s i t y P r e s s ( i npress).
Professor Donald Michie of the University ofEdinburgh includes a descr ip t ion ofHeuristic DENDRAL in his recent paper on“Machines and the Theory of Intelligence”(Nature, 24 1, pp.1973):
507-512, February 23,
“Mass spectrogram an al ysis wasproposed by Lederberg as a suitabletask for machine intelligence methods.The heuristic DENDRAL program
4 2
developed by hirh and Feigenbaumnow outperforms post-doctoralchemists in the identification ofcertain classes of organic compounds.The program is a r ich quarryingground for fundamental mechanismsof intelligence, including thesystematic conjecture of hypotheses,heuristic search, rote learning, anddeductive and inductive reasoning.”
And, in the March, 1973, issue of ComputingReviews (pp. 13% 133), Robert Kling of theUniversity of Wisconsin begins his review ofreference 25 wi th th is assessment ofDENDRAL:
“ T h i s b r i e f p a p e r p r o v i d e s a noverview of o n e of the mostsophisticated applications programs inartificial intelligence.”
Dr. T. G. Evans (Air Force CambridgeResearch Labs), President of the ACMSIGART, in introducing a talk on HeuristicDENDRAL at the 1970 FJCC meeting ofSIGART, called the program “probably thesmartest program in the world” (and followedthis with the interesting observation that thisprogram had probably received a moresustained and intense effort than any othersingle program in the history of the artificialintelligence field).At a practical level, a mass spectrometrylaboratory at the University of Copenhagen,headed by Dr. Gustav Schroll, adapted theprogram to his facilities there..
Bibliography
[I] J. Lederberg, DENDRAL-64 - A Systemfor Computer Construction,Euurneration and Notation of OrganicMolecules as Tree Structures and CyclicGraphs, (technical reports to NASA, alsoavailable from the author andsummarized in [ 121).(la) Part I. Notational algorithm for tree
structures (1964) CR.57029.(1 b) Part II. Topology of cyclic graphs
(1965) CR.68898.
HEURISTIC PROGRAMMING’PROJECT
(lc) Part III. Complete chemical graphs;embedding rings in trees (1969).
[2] J. Lederberg, Computation of MolecularFormulas for Mass Spectrometry, Holden-Day, Inc. ( 1964).
[3] J. Lederberg, Topological Mapping ofOrgarlic Molecules, Proc. Nat. Acad. Sci.,53:1, January 1965, pp, 134-139.
[4l J. Lederberg, Systematics of organicmolecules, graph topology andHamilton circuits. A gerleral outline ofthe DENDRAL system. NASA CR--48899 ( 1965).
[51 J. Lederberg, Harniltoll Circuits ofConvex Trivalent Polyhedra (up to 18vertices), Am. Math. Monthly, May 1967.
[6] G. L. Sutherland, DENDRAL - AComputer Program for Generating andFiltering Chemical Structures, StanfordArtificial Intelligence Memo AIM-49,February 1967.
[7] J. Lederberg and E. A. Feigenbaum,Mechanization of Inductive Infererlcein Organic Chemistry, in B. Kleinmuntz(ed) Formal Representations for HumanJudgment, Wiley, 1968; also StanfordArtificial Intelligence Memo AIM-54,August 1967.
[8] J. Lederberg, Online Computation ofMolecular Formulas from Mass NumberNASA CR-94977, 1968.
191 E. A. Feigenbaum and B. G. Buchanan,Heuristic DENDRAL: A Program forGeneratillg Explauatory Hypotheses inOrganic Chemistry, in Proceedings,Hawaii International Conference onSystem Sciences, B. K. Kinariwala and F,F. Kuo (eds), University of Hawaii Press,1968.
HEURISTIC PROGRAMMING PROJECT 43
[lo] B. G. Buchanan,‘G. L. Sutherland, andE. A. Feigenbaum, HeuristicDENDRAL: A Program for GerleratingExplanatory ‘Hypotheses in OrganicChemistry, in Machine Intelligence 4, B.Meltzer and D. Michie (eds), EdinburghUniversity Press, 1969; also StanfordArtificial Intelligence Memo AIM-62, July1968.
[ 1 l] E. A. Feigenbaum, ArtificialIntelligence: Themes in the SecoudDecade. In Final Supplement toProceedings of the lFlP68 InternationalCongress, Edinburgh, August 1968; alsoStanford Artificial Intelligence MemoAIM-67, August 1968.
[ 121 J. Lederberg, Topology of Molecules, inThe Mathematical Sciences - A Collection0s hays, Committee on Support ofResearch in the Mathematical Sciences(COSRIMS), National Academy ofSciences - National Research Council,
_ M.I.T. Press, 1969, pp. 37-51.
[ 131 C. Sutherland, Heuristic DENDRAL: AFamily of LISP Programs, to appear inD. Bobrow (ed), LlSP Applications; alsoStanford Artificial Intelligence MemoAIM-80, March 1969.
[143 J. Lederberg, C. L. Sutherland, B. G.Buchanan, E. A. Feigenbaum, A. V.Robertson, A. M. Duffield, and C.D jerassi, Applications of ArtificialIntelligence for Chemical Inference I.The Number of Possible OrganicCorn pou uds: Acyclic StructuresContaining C, H, 0 and N. Journal ofthe American Chemical Society, 9 1: I 1, May21, 1969.
[I53 A. M. Duffield, A. V. Robertson, C.D jerassi, B. 6. Buchanan, G. L.Sutherland, E. A. Feigenbaum, and J.Lederberg, AppKcation of ArtificialIntelligence for Chemical Inference II.
Interpretation of Low Resolution MassSpectra of Ketones. Journal of theAmerican Chemical Society, 9 1: 11, May 2 1,1969.
[161 B. G. Buchanan, G. L. Sutherland, E. A.Feigenbaum, Toward an Understandingof Iuformatiou Processes of ScientificInference in the Context of OrganicChemistry, in Machine Intelligence 5, B.Meltzer and D. Michie, (eds), EdinburghUniversity Press 1970; also StanfordArtificial Intelligence Memo AIM-99,September 1969.
El71 J. Lederberg, G. L. Sutherland, B. G.Buchanan, and E. A. Feigenbaum, AHeuristic Program for Solving aScientific Inference Problem: Summaryof Motivation and Implementation,Stanford Artificial Intelligence MemoAIM- 104, November 1969.
[I$] C. W. Churchman and B. G. Buchanan,On the Design of Inductive Systems:Some Philosophical Problems, Britishjournal for the Philosophy of Science, 20,1969, pp. 3 1 l-323.
[191 G. Schroll, A. M. Duffield, C. Djerassi,B. G. Buchanan, G. L. Sutherland, E.A. Feigenbaum, and J. Lederberg,Application of Artificial Intelligencefor Chemical Inference III, AliphaticEthers Diagnosed by Their LowResolution Mass Spectra and NMRData, Journal of the American ChemicalSociety, 91:26, December 17, 1969.
1201 A. Buchs, A. M. Duffield, G. Schroll, C.Djerassi, A. B. Delfino, B. G. Buchanan,G. L. Sutherland, E. A. Feigenbaum, andJ. Lederberg, Applications of ArtificialIutelligeuce For Chemical Inference.IV. Saturated Amines Diagnosed byTheir Low Resolution Mass Spectra andNuclear Magnetic Resonance Spectra,Journal of the American Chemical Society,92,6831, 1970.
44
[21) Y.M. Sheikh, A. Buchs, A.B. Delfino, G.Schroll, A.M. Duffield, C. Djerassi, B.G.Buchanan, G.L. Sutherland, E.A.Feigenbaum and J. Lederberg,Applications of Artificial Iutelligeucefor Chemical Inference V. AnApproach to the Computer Generationof Cyclic Structures. Diff ereutiatiollBetween All the Possible IsotnericKetones of Cotnpositiou C6H100,Organic Mass Spectrometry, 4, 493, 1970.
1221 A. Buchs, A.B. Delfino, A.M. Duffield, C.D jerassi, B.G. Buchanan, E.A.Feigenbaum and J. Lederberg,Applications of Artificial Intelligencefor Chemical Inference VI. Approachto a General Method of IuterpretiugLow Resolution Mass Spectra with aComputer, Chem. Acta Helvetica, 53,1394, 1970.
[23] E.A. Felgenbaum, B.C. Buchanan, andJ. Lederberg, On Generality and
- Problem Solving: A Case Study Usingthe DENDRAL Program, in MachineIntelligence 6, B. Meltzer and D. Michie,(eds.), Edinburgh University Press, 1971;also Stanford Artificial Intelligence MemoAIM- 13 1, August 1970.
1241 A. Buchs, A.B. Delfino, C. Djerassi, A.MDuffield, B.G. Buchanan, E.A.Feigenbaum, J. Lederberg, G. Schroll,and G.L. Sutherland, The Applicationof Artificial IJltelligence in theInterpretation of Low-Resolution MassSpectra, Advances in Mass Spectrometry,5, 314.
[25] B.G. Buchanan and J. Lederberg, TheHeuristic DENDRAL Progranl forExplaining Empirical Data, Proc. /F/PCongress 71; also Stanford ArtificialIntelligence Memo AIM- 14 1.
[26] B.G. Buchanan, E.A. Feigenbaum, andJ. Lederberg, A Heuristic Programming
HEURISTIC PROGRAMMING PROJECT
Study of Theory Formation in Science,Advance Papers of the SecondInternational Joint Conference onArtijcial Intelligence, Imperial College,London, September, 197 1; also StanfordArtificial Intelligence Memo AIM- 145.
[271 Buchanan, B. G., Duffield, A.M.,Robertson, A.V., An Applicatioll ofArtificial Intelligence to theInterpretation of Mass Spectra, MassSpectrometry Techniques and Appliances,George W. A. Milne (ed), John Wiley 8~Sons, 1971, p. 121-77.
[28] D.H. Smith, B.G. Buchanan, R.S.Engelmore, A.M. Duffield, A. Yeo, E.A.Feigenbaum, J. Lederberg, and C.Djerassi, Applications of ArtificialIntelligence for Cheniical InferenceVIII. An approach to the ComputerInterpretation of the High ResolutionMass Spectra of Complex Molecules.Structure Elucidation of EstrogenicSteroids, Journal of the AmericanChemical Society, 94, 5962-597 1, 1972.
[29] B.G. Buchanan, E.A. Feigenbaum, andN.S. Sridharan, Heuristic Theory-Formation: Data Interpretation alldRule Formation, in Machine intelligence7, Edinburgh University Press, 1972.
[31] B. G. Buchanan, Review of HubertDreyfus’ What Computers Can’t Do: ACritique of Artificial Reason, ComputingReviews, January 1973; also StanfordArtificial Intelligence Memo AIM- 18 1,November 1972.
[32] D. H. Smith, B. G. Buchanan, R. S.Engelmore, H. Aldercreutz and C.
Li
-,
HEURISTIC PROGRAMMING PROJECT
D jerassi, Applicat;ons of ArtificialIntelligence for Chemical Inferewe IX.Analysis of Mixtures Without PriorSeparation as Illustrated for Estrogens.Journal of the Amertcan Chemical Society(in press).
1331 D. H. Smith, B. C. Buchanan, W. C.White, E: A. Feigenbaum, C. Djerassiand J. Lederberg, Applications ofArtificial IntelIigence for ChemicalInferewe X. Intsutn. A DataIuterpretatiorl Program as Applied tothe Collected Mass Spectra ofEstrogenic Steroids. Tetrahedron, (inpress).
[34] B. G. Buchanan and N. S. Sridharan,Rule Forlnafion ou Non-HomogeneousClasses of Objects, Advance Papers ofthe Third International Joint Conferenceon Arti..cir,l Intelligence, Stanford,Califorwa, August, 1973.
45
Apphdix A
ACCESS TO DOCUMENTATION
This is a description of how to get copies ofpublications referenced in this report.
Exterrlal Publicatiorls
For books, journal articles, or conferencepapers, first try a technical library. If youhave difficulty, you might try writing theauthor directly, requesting a reprint.Appendix D lists publications alphabeticallyby lead author.
Artificial Intelligence Memos
Artificial Intelligence Memos, which carry an“AIM” prefix on their number, are used toreport on research or development results ofgeneral interest, including all dissertationspublished by the Laboratory. Appendix Blists the titles of dissertations; Appendix Egives the abstracts of all A. I. Memos andinstructions for how to obtain copies. T h etexts of some of these reports are kept in ourdisk file and may be accessed via the ARPANetwork (see below).
Computer Scieme Reports
Computer Science Reports carry a“STAN-CS-” prefix and report researchresults of the Computer Science Department.(All A. I. Memos published since July 1970also carry CS numbers.) To request a copy ofa CS report, write to:
Documentation ServicesComputer Science DepartmentStanford UniversityStanford, California 94306
The Computer Science Department publishesa monthly abstract of forthcoming reportstha t can be reques ted f rom the aboveaddress.
-
47
Film Reports
Several films have been made on researchprojects. See Appendix C for a list of filmsand procedures for borrowing prints.
Operating Notes
Reports that carry a SAILON prefix ( astrained acronym for Stanford A. 1. Lab.Operating Note) are semi-formal descriptionsof programs or equipment in our laboratorythat are thought to be primarily of internalinterest. The texts of most SAILONS areaccessible via the ARPA Network (seebelow). Printed copies may be requestedfrom:
Documentation ServicesArtificial Intelligence Lab.Stanford UniversityStanford, California 94306
Workhg Notes
Much of our working documentation is notstocked in hard copy form, but is maintainedin the computer disk file. Such texts that arein public areas may be accessed from theARPA Network (see below). Otherwise,copies may be requested from the author(s) ’at the address given above for OperatingNotes.
Public File Areas
People who have access to the A R P ANetwork are welcome to access our publicfiles. The areas of principal interest andtheir contents are a follows:
[BIB,DOC] bibliographies of variouskinds,
[AIM,DOC] texts of a few of our A. I.Memos,
[S,DOCl many of our SAILONs,[UP,DOCl working notes (quite informal),[P,DOCl “people-oriented” files, including
the lab phone directory.
48
Network Access
To get into our system from the Network, say"L NET .GUE", which logs you in as a “networkguest”. All commands end with a carriagereturn. Our monitor types ‘I.” whenever it isready to accept a command. To halt aprogram that is running, type <Control&twice.
If your terminal has both upper and lowercase characters, let the monitor know bySaying ))TTY FULLI) . If you are at a typewriterterminal; you may also wish to type "TTY FILL",
which causes extra carriage returns to beinserted so that the carriage has time toreturn to the left margin before the next linebegins.
To see the names of all the files in, say, the[S,DOC] area (where SAILONs are stored),type "DIR tS,DOCl ' . This will produce a list offiles and the dates they were last written.Among others, it should list “INTRO.TES”,which is an introduction to our timesharingsystem (usually obsolescent, alas), written bythe programmer whose initials are TES.
To type out the contents of a given file, suchas INTRO.TES, say
TYPE INTRO .TEStS ,OOCland it will come spewing forth. To stop thetypout, say <Control>C twice and it will stopafter a few lines. To continue, type "CONT". Ifyou wish to examine selected portions of textfiles, use the SOS editor in read-only mode,as described in SOS.LES[S,DOCl. .
To log out when you are done, typeK <carriage return>
There may be some difficulty with files thatemploy the full Stanford character set, whichuses some 26 of the ASCII control codes (0 to37 octal) to represent special characters.
ACCESS TO DOCUMENTATION
File Transfer
Files can also be transferred to another siteusing the File Transfer Protocol.Documentation on our FTP program i slocated disk file i nFTP.DCS&i~,DO~~r N o p a s s w o r d s o raccount numbers are needed to access ourFTP from the outside.
-
App&dix B
THESES
i.
Theses that have been published by theStanford Artificial Intelligence Laboratoryare listed here. Several earned degrees atinstitutions other than Stanford, as noted.
.c D. Raj. Reddy, AIM-43
I An Approach to Computer SpeechRecognition by Direct Analysis of theSpeech Wave,Ph.D. Tliesis in Computer Science,September 1966.
S. Persson, AIM-46Some Sequence Extrapolating Programs: aStudy of Represeutatioll and Modeling inIiiquiring Systems,Ph.D. Thesis in Computer Science,University of California, Berkeley,September 1966.
Bruce Buchanan, AIM-47Logics of Scientific Discovery,Ph.D. Thesis in Philosophy, University ofCalifornia, Berkeley,December 1966.
James Painter, AIM-44Semantic Correctness of a Compiler for anA lgol-like Language,Ph.D. Thesis in Computer Science,March 1967.
,
William W ichman, AIM-56IJse of Optical Feedback in the ComputerControl of an Arm,Eng. Thesis in Electrical Engineering,August 1967.
Monte Cailero, AIM-58An Adaptive Command and ControlSystem Utilizing Heuristic LearningP recesses,Ph.D. Thesis in Operations Research,December 1967.
49
Donald Kaplan, AIM-60The Formal Theoretic Allalysis of SttolrgEquivalence for Elemental Properties,Ph.D. Thesis in Computer Science,July 1968.
Barbara Huberman, AIM-65A Program to Play Chess End Games,Ph.D. Thesis in Computer Science,August 1968.
Donald Pieper, AIM-72The Kinematics of Manipulators underComputer Control, :Ph.D. Thesis in Mechanical Engineering,October 1968.
Donald Waterman, AIM-74Machine Learning of Heuristics,Ph.D. Thesis in Computer Science,December 1968.
Roger Schank, AIM-83A Conceptual Dependemy Representationfor a Computer Oriented Semantics,Ph.D. Thesis in Linguistics, University ofTexas,March 1969.
Pierre Vicens, AIM-85 ’Aspects of Speech Recognition byComputer,Ph.D. Thesis in Computer Science,March 1969.
Victor D. Scheinman,Desigll of Computer Controlled-Manipulator,
AIM-92
Eng. Thesis in Mechanical Engineering,June 1969.
Claude Cordell Green, AIM-96The Application of Theorem Proving toQuestion-answering System,Ph.D. Thesis in Electrical Engineering,August 1969.
THESES
James J. Horning, . AIM-98A Study of Grammatical Inference,Ph.D. Thesis in Computer Science,August 1969.
Michael E. Kahn, AIM- 106The Near-minimum-time Control of Open-loop Articulated Kinematic Chains,Ph.D. Thesis in Mechanical Engineering,December 1969.
Irwin Sobel, AIM-121Camera Models and Machine Perception,Ph.D. Thesis in Electrical Engineering,May 1970.
Michael D. Kelly, AIM- 130Visual Identification of People byComputer,Ph.D. Thesis in Computer Science,July 1970.
Gilbert Falk, AIM-132Computer Interpretatiorl of Imperfect LineData as a Three-dimensional Scene,Ph.D. Thesis in Electrical Engineering,August 1970.
Jay Martin Tenenbaum, AIM- 134Accommodation in Computer Vision,Ph.D. Thesis in Electrical Engineering,September 1970.
Lynn H. Quam, AIM- 144Corn pu ter Com parisou of Pictures,Ph.D. Thesis in Computer Science, _May 1971.
Robert E. Kling, AIM- 147Reasoning by Analogy with Applicatiorlsto Heuristic Problem Solving: a Case Study,Ph.D. Thesis in Computer Science,August- 1971.
Rodney Albert Schmidt, AIM- 149A Study of the Real-time Control of aCorn puter-driven Vehicle,Ph.D. Thesis in Electrical Engineering,August 1971.
Jonathan Leonard Ryder, AIM-155Heuristic Analysis of Large Trees asGenerated in the Game of Co,Ph.D. Thesis in Computer Science,December 197 1.
Jean M. Cadiou, AIM- 163Recursive Definitions of Partial Functionsand their Computations,Ph.D. Thesis in Computer Science,April 1972.
Gerald Jacob Agin, AIM-173Representation and Description of CurvedObjects,Ph.D. Thesis in Computer Science,October 1972.
Francis Lockwood Morris, AIM- 174Correctness of Translqtions ofProgramming Languages -- an AlgebraicApproach,Ph.D. Thesis in Computer Science,August 1972.
Richard Paul, AIM- 177Modelling, Trajectory Calculation andServoing of a Computer Controlled Arm,Ph.D. Thesis in Computer Science,November 1972.
Aharon Gill, AIM-178Visual Feedback and Related Problems inComputer Controlled Hand EyeCoordination,Ph.D. Thesis in Electrical Engineering,October 1972.
Ruzena Bajcsy, AIM- 180Computer Ideutificatiorr of TexturedVisiual Scenes,Ph.D. Thesis in Computer Science,October 1972.
Ashok Chandra, AIM- 188On the Properties and Applications ofProgramming Schemas,Ph.D. Thesis in Computer Science,March 1973.
-
THESES 51
Gunnar Rutgcr Grape, AIM&$04Model Based (Intermediate Level) ComputerVisiotr,Ph.D. Thesis in Computer Science,May 1973.
Yoram Y aklmovsky, AIM-209Sctnt Analysis Using a Semantic Base farRegion Growing,Ph.D. Thesis in Computer Science,July 1973.
Appendix C 3. Raj Reddy, Dave Espar and Art Eisenson,
FILM REPORTSHear Here, 16mm color with sound, 15minutes, March 1969.
.Prints of the following films are available forshort-term loan to interested groups without
I charge. They may be shown only to groupsI that have paid no admission fee. To make a
. reservation, write to:L Film Services
Artificial Intelligence Lab.Stanford UniversityStanford, California 94305
Alternatively, prints may be purchased atcost (typically $30 to $50) from:
Cine-Chrome Laboratories4075 Transport St.Palo Alto, California(415) 321-5678
1. Art Eisenson and Gary Feldman, Ellis D.Kroptechev and Zeus, his Marvelous
-Time-sharing System, 16mm B&W withsound, 15 minutes, March 1967.
The advantages of t ime-shar ing overstandard batch processing are revealedthrough the good offices of the Zeus time-sharing system on a PDP-1 computer. Ourhero, Ellis, is saved from a fate worse thandeath. Recommended for mature audiencesoilly.
2. Gary Feldman, Butterfinger, 16mm colorwith sound, 8 minutes, March 1968. .
c
Describes the state of the hand-eye system atthe Artificial Intelligence Project in the fall of1967. The PDP-6 computer getting visualinformation from a television camera andcon trolling a n electrical-mechanical armsolves simple tasks involving stacking blocks.The techniques of recognizing the blocks andtheir positions as well as controlling the armare briefly presented. Rated (3.
-
53
Describes the state of the speech recognitionproject as of Spring, 1969. A discussion ofthe problems of speech recognition isfollowed by two real time demonstrations ofthe current system. The first shows t h ecomputer learning to recognize phrases andsecond shows how the hand-eye system maybe con trolled by voice commands.Commands as complicated as ‘Pick up thesmall block in the lower lefthand corner’, arerecognized and the tasks are carried out bythe computer controlled arm.
4. Gary Feldman and Donald Peiper, Avoid,16mm silent, color, 5 minutes, March1969.
Reports on a computer program written byD . P e i p e r f o r h i s P h . D . T h e s i s . T h eproblem is to move the computer controlledelectro-mechanical arm through a space filledwith one or more known obstacles. T h eprogram uses heuristics for finding a safepath; the film demonstrates the arm as itmoves through various clutteredenvironments with fairly good success.
5. Richard Paul and Karl Pingle, InstantInsanity, 16mm color, silent, 6 minutes,August, 1971.
Shows the hand/eye system solving thepuzzle Instant Insanity. Sequences include%nding a n d recognizing cubes, colorrecognition and object manipulation. Thisfilm was made to accompany a paperpresented at the 1971 International JointConference o n Artificial Intelligence inLondon and may be hard to understandwithout a narrator.
A technical presentation of three researchprojects completed in 1972: advanced armcontrol by R. P. Paul [AIM-1773, v isualfeedback control by A. Gill [AIM-1781, andrepresentation and description of curvedobjects by G. Agin [AIM- 1733.
7. Larry Ward, Computer InteractivePicture Processing, (MARS Project),16mfn color, sound, 8 min., Fall 1972.
This film describes an automated picturedifferencing technique for analyzing thevariable surface features on Mars using datareturned by the Mariner 9 spacecraft. Thesystem uses a time-shared, terminal orientedPDP- 10 computer. The film proceeds at abreath less pace. Don’t blink, or you will missan entire scene.
8.- Richard Paul and Karl Pingle,Automated Pump Assembly, 16mmcolor, silent (runs at sound speed!), 7minutes, April, 1973.
Shows the hand-eye system assembling asimple pump, using vision to locate the pumpbody and to check for errors. The parts areassembled and screws inserted, using somes’pecial tools designed for the arm. Sometitles are included to help explain the film.
9. Terry. Winograd, Dialog with a robot,16mm black and white, silent, 20 minutes,(made at MIT), 1971.
Presents a natural language dialog with asimulated robot block-manipulation system.The dialog is substantially the same as thatin Understanding Natural Language (T .Winograd, Academic Press, 1972). Noexptanatory or narrative material is on thefilm.
FILM REPORTS
Appekdix D
EXTERNAL PUBLICATIONS
Articles and books by Project members arelisted here alphabetically by lead author.Only publications following the individual’saffiliation with the Project are given.
1. Agin, Gerald J., Thomas 0. Binford,Computer Description of CurvedObjects, Proceedings of the ThirdInternational Joint Conference onArtijcial! Intelligence, StanfordUniversity, August 1973.
2. Allen, John, David Luckham, AnIllteractive Theorem-Proving Programin Bernard Meltzer and Donald Miqhie(eds.), Machine Intelligence 5, EdinburghUniversity Press, 1970.
3. Ashcroft, Edward, Zohar Manna,Formalizatiorl of Properties of ParallelPrograms, Machine Intelligence 6,Edinburgh Univ. Press, 1971.
4. Ashcroft, Edward, Zohar Manna, TheTranslation of ‘Go To’ Programs to‘While’ Programs, Proc. /F/P Congress1571.
5. Ashcroft, Edward, Zohar Manna, AmirPnueli, Decidable Properties of MonadicFunctional Schemas, J. ACM, July 1973.
6. Bajcsy, Ruzena, Computer Descriptionof Textured Scenes, Proc. Third ht.Joint Con.. on Arti@al Intelligence,Stanford U., 1973.
7. Beauchamp, James, H. Van Foerster(eds.),-Music by Computers, John Wiley,New York, 1969.
-
55
8. Becker, Joseph, The Modeling of SimpleAllalogic arid Inductive Processes in aSemantic Memory System, Proc.International Con.. on ArtijicialIntelligence, Washington, D.C., 1969.
9. Biermann, Alan, Jerome Feldman, On theSynthesis of Finite-state Machines fromSamples of Their Behavior, IEEETransactions on Computers, Vol. C-21,No. 6, pp. 592-596, June 1972.
10. Biermann, Alan, On the Inference ofTurirlg Machines from SampleComputations, Arti$cial intelligence J.,-Vol. 3, No. 3, Fall 1972.
11. Binford, Thomas O., Sensor Systemsfor Manipulation, in E. Heer (Ed.),Remotely Manned Systems, Calif. Inst. ofTechnology, 1973.
12. Binford, Thomas, Jay M. Tenenbaum,Computer Vision, Computer (IEEE),May 1973.
13. Bracci, Ciampio, Marco Somalvico, AJIInteractive Software System forComputer-aided Design: An Applicationto Circuit Project, Comm. ACM,September 1970.
14: Buchanan, Bruce, Georgia Sutherland,Heuristic Dendral: A Program forGenerating Hypotheses in OrganicChemistry, in Donald Michie (ed.),Machine Intelligence 4, AmericanElsevier, New York, 1969.
15. Buchanan, Bruce, Georgia Sutherland,Edward Feigenbaum, Rediscoverillgsome Problems of Artificial Intelligencein the Corltext of Organic Chemistry,in Bernard Meltzer and Donald Michie(eds), Machine Intelligence 5, Edin burghUniversity Press, 1970.
56 EXTERNAL PUBLICATIONS
16. Buchanan, Bruce, T. Headrick, SomeSpeculation about Artificial Intelligenceand Legal Reasoning, Stanford LawRezhw, November 1970.
17. Buchanan, Bruce, A. M. Duffield, A. V.Robertson, An Application of ArtificialIntelligence to the Interpretation ofMass Spectra, in Mass Spectromet~yTechniques and Appliances, George W.Milne (ed), John Wiley & Sons, 1971.
18. Buchanan, Bruce, Edward Feigenbaum,Joshua Lederberg, A HeuristicProgramming Study of TheoryFormation in Science, Proc. SecondInternational Joint Conference on ArijicialIntelligence (21JCAI), British ComputerSociety, Sept. 197 1.
19. Buchanan, Bruce, Joshua Lederberg,The Heuristic DENDRAL Program forExplaining Empirical Data, Proc. /F/PCongress 191.
20. Buchanan, Bruce, E. A. Feigenbaum,and N. S. Sridharan, Heuristic TheoryFormation: Data Interpretation andRule Formation, in Madine intelligence7, Edinburgh University Press, 1972.
21. Buchanan, Bruce C., Review of HubertDreyfus’ ‘What Computers Can’t Do’: ACritique of Artificial Reason, ComputingReview, January 1973.
22. Buchanan, Bruce, N. S. Sridharan,Analysis of Behavior of ChemicalMolecules: Rule Formation OJI Non-Homogeneous Classes of Objects,Proceedings of the Third InternationalJoint Conference on Artificial intelligence,Stanford University, August 1973.
23. Buchs, A., A. Delfino, A. Duffield, C.Djerassi, B. Buchanan, E. Feigenbaum,J. Lederberg, Applicatious of ArtificialIutelligerlce for Chemical Inference VI.
Approach to a General Method ofIilterpreting Low Resolution MassSpectra with a Computer, HelveticaChemica Acta, 53:6, 1970.
24. Buchs, A., A. Duffield, G. Schroll, CarlDjerassi, A. Delfino, Bruce Buchanan,Georgia Sutherland, EdwardFeigenbaum, Joshua Lederberg,Applications of Artificial Intelligencefor Chemical Inference IV. SaturatedAmiues Diagnosed by their LowResolutioll Mass Spectra and NuclearMaglletic Resonance Spectra, J. Amer.
_ Chem. Sot., 92:23, November 1970.
25. Cadiou, Jean M., Zohar Manna,Recursive Definitious of PartialhliCtiOllS and their Computations,ACM SIGPLAN Notices, Vol. 7, No. 1,January 1972.
26. Campbell, John, AlgebraicComputatiou of Radiative Correctionsfor Electron-Proton Scattering, NuclearPhysics, Vol. B I, pp. 238-300, 1967.
27. Campbell, Anthony Hearn and J. A.,Symbolic Analysis of FeynmauDiagrams by Computer, Journal ofComputational Physics 5, 280-327, 1970.
28. Chowning, John M., The Simulation ofMoving Sound Sources, Proc. AudioEngineering Sot. Convention, May 1970.
29. Chowning, John M., The Synthesis ofComplex Audio Spectra by means ofFrequency Modulation, J. AudioEngineering Society, September 1973.
30. Churchman, C., Bruce Buchanan, Onthe Desiglt of Inductive Systems: SomePhilosophical Problems, British Journalfor the Philosophy of Science, 20, 1969, pp.31 l-323.
EXTERNAL PIJBLICATIONS 57
31. Colby, Kenneth, David Smith, Dialoguesbetween Humails and Artificial BeliefSystems, Proc. International Conferenceon Artificial Intelligence, Washington,D.C., 1969.
32. Colby, Kenneth, Larry Tesler, HoraceEnea, Experiments with a SearchAlgorithm for the Data Base of aHuman Belief System, Proc.International Conference on ArtificialIntelligence, Washington, D.C., 1969.
33. Colby, Kenneth Mark, The Rationalefor Computer-Based Treatment ofLanguage Difficulties in NonspeakingAutistic Children, Journal of Autism andChildhood Schizophrenia, Vol. 3, 254-260,1970.
34. Colby, Kenneth, Mind and Braili Again,Mathematical Biosciences, Vol. 11, 47-52,1970.
36. Colby, Kenneth, F. Hilf, S. Weber, H. C.Kraemer, Turing-likeIndistinguishability Tests for theValidatiou of a Computer Simulationof Paranoid Processes, ArtijcialIntelligence J., Vol. 3, No. 3, Fall 1972.
37. Colby, Kenneth M., The Ratiotiale forCoin pu ter-based Treatmerit ofLanguage Difficulties in NonspeakingAutistic Children, Journal of Autism andChildhood Schizophrenia, Vol. 3, 254-260,1973.
38. Dobrotin, Boris M., Victor D.Scheinman, Design of a ComputerC&trolled Manipulator for RobotResearch, Proc. Third ht. Joint Conf. onArtijcial Intelligence, Stanford U., 1973.
39. Duffield, Alan, A. Robertson, CarlDjerassi, Bruce Buchanan, G.Sutherland, Edward Feigenbaum, JoshuaLederberg, Application of ArtificialIntelligence for Chemical InterferenceII. Irlterpretatiou of Low ResolutionMass Spectra of Ketones, J. Amer.CAem. Sot., 9 1: 1 1, May 1969.
40. Enea, Horace, Kenneth Mark Colby,Idiolectic Language-Analysis forUnderstatiding Doctor-PatientDialogues, Proceedings of the ThirdInternational Joint Conference onArtificial Intelligence, StanfordUniversity, August 1973.
41. Falk, Gilbert, Scelle Analysis Based onImperfect Edge Data, Proc. 2IJcAI,Brit. Comp. Sot., Sept. 1971.
42. Falk, Gilbert, Interpretation ofImperfect Line Data as a Three-dltnensioual Scene, Artificial IntelligenceJ., Vol. 3, No. 2, 1972.
43. Feigenbaum, Edward, InformationProcessirlg and Memory, in Proc. FifthBerkeley Symposium on MathematicalStatistics and Probability, Vol. 4, U.C.Press, Berkeley, 1967.
44. Feigenbaum, Edward, Joshua Lederberg,Bruce Buchanan, Heuristic Dendral,Proc. International Conference on SystemSciences, University of Hawaii and IEEE,University of Hawaii Press, 1968.
45. Feigenbaum, Edward, ArtificialIntelligence: Themes in the SecondDecade, Proc. IF IP Congress, 1968.
46. Feigenbaum, Edward, Bruce Buchanan,Joshua Lederberg, On Generality andProblem Solving: A Case Study usingthe DENDRAL Program, MachineIntelligence 6, Edinburgh Univ. Press,1971.
58
47. Feldman, Jerome, D. Cries, TraklatorWriting Systems, Comm. ACM, February1968.
48. Feldman, Jerome, P. Rovner, The LeapLanguage Data Structure, Pm. lFIPCongt P55, 1968.
49. Feldman, Jerome, Machine Iutelligence,review of Numbers I-III of the MachineIntelligence series, information andControl, 14, 490-492, 1969.
50. Feldman, Jerome, Towards AutomaticProgram In iug, Preprints of NATOSqftruure Engineering Conference, Rome,Italy, 1969. *
5 1. Feldman, Jerome, Paul Rovner, AnAlgol-based Associative Language,Comm. ACM, August 1969.
52. Feldman, Jerome, Gary Feldman, G.Falk, Gunnar Grape, J. Pearlman, I.
- Sobel, and J. Tenenbaum, The StaufordHand-Eye Project, PYOC. internationalConf. on Artificial Intelligence,Washington, DC., 1969.
53. Feldman, Jerome, Getting a Computerto See Simple Scenes, IEEE StudentJournal, Sept. 1970.
54. Feldman, Jerome, Alan Bierman, ASurvey of Grammatical Inference, Proc.International Congress on PatternRecognition, Honolulu, January 1971, alsom S, Watanbe (ed.), Frontiers of PatternRecognition, Academic Press, 1972.
55. Feldman, Jerome, Robert Sproull,System Support for the Stanford Hand-eye System, Proc. 21JCAI, Brit. Comp.Sot., Sept. 1971.
56. Feldman, Jerome, et al, The Use ofVisiorl arid Mauipulatiou to Solve the‘Instant Insanity Puzzle, Proc. 21JCAI,Brat. Comp. Sot., Sept. 1971.
EXTERNAL PIJRLICATTONS
57. Feldman, Jerome, Sonle DecidabilityResults 011 Gramnlatical Inference andComplexity, Information and Control,Vol. 20, No. 3, pp 243-262, April 1972.
58. Feldman, Jerome, J. Low, D. Swinehart,R. Taylor, Recent Developments inSAIL, an ALGOL-based language forArtificial Intelligence, Proc. Fall JointComputer Conference, 1972.
59. Floyd, Robert, Toward InteractiveDesign of Correct Programs, Proc. IFIPCongress 1571.
60. Garland, Stephan J., David Luckham,Translatiilg Recursive Schelnes intoProgram Schemes, ACM SKPLANNotxes, Vol. 7, No. I, Jarluary 1972.
61. Garland, Stephan J., David C. Luckham,Program Schemes, Recursiou Schemes,and Formal Languages, J (‘ntrlfJ7/rer andSystem Sciences, Vol. 7, No. 2, April 1973.
62. Gips, James, A New Reversible Figure,Perceptual U Motor Skills, 34, 306, 1972.
63. Goodman, Joseph, Digital ImageFormatioll from Electronically DetectedHolograms, Applied Physics Letters,August 1967.
64. Goodman, Joseph, Digital ImageFomatiou from Electronically DetectedHolograms, in Proc. SPIE Seminar onDtgitnl Imaging Techniques, Sot. Photo-Optical Instrumentation Engineering,Redondo Beach, California, 1967.
65. Gruen, Richard, William Welher, RapidProgram Generation, Proc. LIEECUSSymposium, Fall 1968.
66. Hearn, Anthony, Computatiorl ofAlgebraic Properties of ElementaryParticle Reactions Using a DigitalComputer, Comm. ACM, 9, pp. 573-577,August, 1966.
EXTERNAL PUBLICATIONS
67. Hearn, Anthony, REDUCE, A User-Oriented Interactive System forAlgebraic Simplification, Proc. ACMSymposium on Interactive Systems forExper imentai Applied Mathematics,August 1967.
68. Hearn, Anthony, The Problem ofSubstitution, Proc. IBM SummerInstitute on Symbolic Mathematics byComputer, July 1968.
69. Heal-n, Anthony, Applications ofSymbol Marlipulatiorl in TheoreticalPhysics, Comm. ACM, August 1971.
70. Hilf, Franklin, Kenneth Colby, DavidSmith, W. Wittner, William Hall,Machine-Mediated Interviewing, J.Nervous 6’ Mental Disease, Vol. 152, No.4, 1971.
7 1. Hilf, Franklin, Non-NonverbalCommuuicatiou and PsychiatricResearch, Archives of General Psychiatry,
_ Vol. 27, November 1972.
77. Kaplan, Donald, Some CompletenessResults in the Mathematical Theory ofComputation, ACM Journal, January1968.
78. Kaplan, Donald, Regular Expressionsand the Completeness of Programs, J.Camp. &? System Sci., Vol. 3, No. 4, 1969.
79. Katz, Shmuel, Zohar Manna, AHeuristic Approach to ProgramVerification, Proceedings of the ThirdInternational Joint Conference onArtificial Intelligence, StanfordUniversity, August 1973.
80. Kieburtz, Richard B., David Luckham,Compatibility and Complexity ofRefinements of the ResolutionPrincipal, SIAM J. Compuf., 1972.
81. Kieburtt, Richard, David Luckham,Com patability and Complexity ofRefinemeuts of the ResolutionPrinciple, SIAM J. on Computing, 1-4,1973.
72. Hilf, Franklin, Partially AutomatedPsychiatric Research Tool, J. Nervousand Mental Disease, Vol. 155, No. 6,December 1972.
82. Kling, Robert, A Paradigm forReasoning by Analogy, Proc. 21JCAI,Brit. Comp. Sot., Sept. 1971.
73. Hueckel, Manfred, An Operator whichLocates Edges iu Digitized Pictures,JACM, January 1971.
83. Knuth, Donald E., The Art of ComputerProgramming, Vol. 2, SeminumericalAlgorithms, Addison-Wesley, Menlo Park,Calif., 1969.
74. Hueckel, Manfred H., A Local VisualOperator which Recognizes Edges andLines, J. ACM, October 1973.
84. Knuth, Donald E., An Empirical Study- of FORTRAN Programs, Software --
Practice and Experience, Vol. 1, 105- 133,1971.
75. Ito, T., Note 011 a Class of StatisticalRecognition Functions, IEEE Trans.Computers, January 1969.
85. Knuth, Donald E., Ancient BabylonianAlgorithms, Comm. ACM, July 1972.
76. Kahn, Michael, Bernard Roth, The 86. Knuth, Donald E., The Art of ComputerNear-minimum-time Coritrol of Open- Programming, Vol. 3, Sorting andloop Articulated Kinematic Chains, Searcldng, Addison-Wesley, Menlo Park,Trans. ASME, Sept. 1971. Calif., 1973.
88. Lederberg, Joshua, Edward Feigenbaum,Mechanization of Inductive Iufereuceirr Orgarlic Chemistry, in B. Kleinmuntz(ed.), Formal Representation of HumanJudgment, John Wiley, New York, 1968.
89. Lederberg, Joshua, Topology ofOrganic Molecules, National Academy ofScience, The Mathematical Sciences: aCollection of Essays, MIT Press,Cambridge, 1969.
90. Lederberg, Joshua, Georgia Sutherland,Bruce Buchanan, Edward Feigenbaum,A. Robertson, A. Duffield, Carl Djerassi,Applications of Artificial Iiitelligetlcefor Chemical lrlferewe I. The Numberof Possible Organic Compounds:Acyclic Structures Containing C, H, 0,
- aiid N, J. Amer. CAem. Sot., 91:11, May1969.
9 1. Lederberg, Joshua, Georgia Sutherland,Bruce Buchanan, Edward Feigenbaum, AHeuristic Program for Solving aScientific Iufereuce Problem: Summaryof Motivatioll and Implementation, inM. Mesarovic (ed.), TheoreticalApproaches to Non-numerical ProblemSolving, Springer-Verlag, New York,1970.
92. London, Ralph, Correctness of aCompiler for a LISP Subset, ACMSIGPLAN Notices, Vol. 7, No. 1,January 1972.
93. Luck-ham, David, Refinement Theoremsin ~Resolution Theory, Proc. I%8 /R/ASymposium in Automatic Deduction,Versailles, France, Springer-Verlag, 1970.
94. Luckham, David, D. Park and M.Paterson, On Fortnalised ComputerPrograms, J. Camp. U System Sci., Vol.4, No. 3, June 1970.
95. Luckham, David, Nils Nilsson,Extracting Iuformatiou fromResolutioii Proof Trees, ArtificialIntelligence Journal, Vol. 2, No. 1, pp.27-54,. June 1971.
96. Luckham, David C., AutomaticProblem Solving, Proceedings of theThird International Joint Conference onArtijcial Inteliigence, Stanford-University, August 1973.
97. Manna, Zohar, Properties of Programsarrd the First Order Predicate Calculus,J. ACM, Vol. 16, No. 2, April 1969.
98. Manna, Zohar, The Correctness ofPrograms, J, System and ComputerSciences, Vol. 3, No. 2, May 1969.
99. Manna, Zohar, John McCarthy,Properties of Programs aud PartialFuiictioii Logic in Bernard Meltzer andDonald Michie (eds.), MachineIntelligence 5, Edinburgh UniversityPress, 1970.
101. Manna, Zohar, ?‘eriiiiliatiori ofAlgorithms Represented as InterpretedGraphs, AF IPS Conference Proc. (SJCC),Vol. 36, 1970.
102. Manna, Zohar, Second-orderMathematical Theory of Computation,Proc. ACM Symposium on Theory ofComputing, May 1970.
EXTERNAL PUBLICATIONS
103. Manna, Zohar, Amir Pnueli,Formalization of Properties ofFwlctional Programs, J. ACM, Vol. 17,No. 3, Juty 1970.
104. Manna, Zohar, R. Waldinger, TowardAutomatic Program Synthesis, Comm.ACM, March 1971.
105, Manna, Zohat-, Mathematical Theoryof Partial Correctness, J. Comp. U Sgs.Sci., June 1971.
106. Manna, Zohar, S. Ness, J. Vuillemin,Inductive Methods for ProvillgProperties of Prograins, ACMSIC; P LA N Notices, Vol. 7, No. 4,January 1972.
107. Manna, Zohar, J. Vuillemin, FixpoiutApproach to the Theory ofComputation, Comm. ACM, July 1972.
108. Manna, Zohar, Program Schelnas, inCurrents in the Theory of Computing (A.V. Aho, Ed.), Prentice-Hall, EnglewoodCliffs, N. J., 1973.
109. Manna, Zohar, Stephen Ness, JeanVuillemin, Inductive Methods forProving Properties of Programs, Comm.ACM, August 1973.
110. Manna, Zohar, AutotnaticProgramln ing, Proceedings of the ThirdInternational Joint Conference onArtijcial intelligence, Stanford .University, August 1973.
1 1 1. Manna, Zohar, Introduction toMathematical Theory of Computation,McGraw-Hill, New York, 1974.
112. McCarthy, John, Towards aMathematical Theory of Computation,in Proc. IF/P Congress 62, North-Holland, Amsterdam, 1963.
61
113. McCarthy, John, A Basis for aMathematical Theory of Computation,in P. Biaffore and D. Hershberg (eds.),Computer Programming and FormalSystems, North-Holland, Amsterdam,1963.
114. McCarthy, John, S. Boilen, E. Fredkin,J.C.R. Licklider, A Time-SharingDebugging System for a SmallComputer, PYOC. A F IPS Conf. (S JCC),Vol. 23, 1963.
115. McCarthy, John, F, Corbato, M.Daggett, The Linking SegmentSubprogram Language and LinkingLoader Program lning Languages, Comm.ACM, July 1963.
116. McCarthy, John, Problems in theTheory of Computation, Proc. IF/PCongress 1565.
117. McCarthy, John, Time-SharingComputer Systems, in W. Orr (ed.),Conversational Computers, Wiley, 1966.
118. McCarthy, John, A FormalDescription of a Subset of Algol, in T.Steele (ed.), Formal Language DescriptionLanguages for Computer Programming,North-Holland, Amsterdam, 1966.
119. McCarthy, John, Infortnation,Scienti.c American, September 1966.
120. McCarthy, John, Computer Control ofa Hand aud Eye, in Proc. Third AU-Union Conference on Automatic Control(Technical Cybernetics), Nauka, Moscow,1967 (Russian).
121. McCarthy, John, D. Brian, G. Feldman,and J. Allen, THOR -- A Display BasedTime-Sharing System, Proc. AFIPSConf. (FJCC), Vol. 30, Thompson,Washington, DC., 1967.
62 EXTERNAL PUBLICATIONS
122. McCarthy, John, James Painter,Correctness of a Conipiler forArithmetic Expressions, Amer. Math.Sot., Proc. Symposia in Applied Math.,Math. Aspects of Computer Science, NewYork, 1967. .
131. Montanari, Ugo, On Limit Propertiesin Digitization Schemes, JACM, April1970.
132. Moncanari, Ugo, Separable Crap hs,Planar Graphs and Web Grammars,Information and Control, M ay 1970.
123. McCarthy, John, Programs withCommon Sense, in Marvin Mir\sky (ed.),Semantic Information Processing, MITPress, Cambridge, 1968.
124. McCarthy, John, Lester Earnest, D.Raj: Reddy, Pierre Vicens, A Computerwith Hands, Eyes, and Ears, Proc.AFIPS Conf. (FJCC), 1968.
125. McCarthy, John, Patrick Hayes, SomePhilosophical Problems from theS tartdpoin t of Artificial Iutelligeuce, inDonald Michie (ed.), Machine Intelligence4, American Elsevier, New York, 1969.
l-26. Milner, Robin, An AlgebraicDefirtitiou of Silnulatiou betweeuPrograms, Proc. 21JCAI, Brit. Comp.Sot., Sept. 197 1.
127. Milner, Robin, Implelnentatiion andApplicatiorl of Scott’s Logic forColnputable Functions, ACM SIGPLAPINOTICES, Vol. 7, No. 1, January 1972.
128. Milner, Robin, Richard Weyhrauch,Proviiig Compiler Correctness in aMechanized Logic, Machine Intelligence7, Edinburgh University Press, 1972.
130. Montanari, Ugo, A Note on MinimalLength Polygonal Approximation to aDigitized Contour, Comm. ACM, January1970.
134. Montanari, Ugo, On the Optimal- Detectioll of Curves ill Noisy Pictures,
Comm. ACM, May 1971.
135. Moorer, James A., Dual Processing forthe PDP-6/10, Decuscope, Vol. 8, No. 3,1969.
136. Moorer, James A., Music andComputer Composition, Comm. ACM,January 1972.
137. Nevatia, Ramakant, Thomas 0.Binford, Structured Descriptions ofComplex Objects, Proceedings of theThird International Joint Conference onArtificial Intelltgence, Stan fordUniversity, August 1973.
138. Nilsson, Nils, Problem-solving Methodsin Artijciat Intellegence, McGraw-Hill,New York, 1971.
139. Paul, Richard, G. Falk, JeromeFeldman, The Computer Representationof Simply Described Scenes, Proc. 2ndIllinois Graphics Conference, Univ.Illinois, April 1969.
140. Paul, Richard, Gilbert Falk, JeromeFeldman, The Computer Description ofSimply Described Scertes, in PertinentConcepts in Computer Graphics, J.Neivergelt and M. Faiman (eds.), U.Illinois Press, 1969.
EXTERNAL PUBLICATIONS
141. Paul, Richard, Trajectory Control of aComputer ArJn, Proc. 2IJCAI, Brit.Comp. Sot., Sept. 197 1.
142. Pingle, Karl, J. Singer, and W.Wichman, Computer CoJltrol of aMechanical Arm through Visual Input,Proc. IFIP Congress 1968, 1968.
143. Pingle, Karl, Visual Perception by aCoJnputer, Automatic Interpretation andClassijcation of Images, Academic Press,New York, 1970.
145. O,LIam, Lynn, Robert Tucker, BotondEross, J. Veverka and Carl Sagan,Mariner 9 Picture DiffereJlciJlg atStallford, S&y and Telescope, August1973.
146. Reddy, D. Raj., SegJneJJtatioJl of- Speech SoJJuds, J. Acoust. Sot. Amer.,
August 1966.
147. Reddy, D. Raj., PhoneJne GroupiJlg forSpeech Recognition, 1. Acoust. sot.Amer., May 1967.
148. Reddy, D. Raj., Pitch PeriodDetermination of Speech Sounds, Comm.ACM, June 1967.
149. Reddy, D. Raj., Computer Recognitiorlof Conrlected Speech, J. Acoust. Sot.Amer., August 1967.
150. Reddy, D. Raj., ComputerTraJlscriptioJr of Phonemic Sytnbols, J.Acoust. Sot. Amer., August 1968.
15 1. Reddy, D. Raj., ConsonaJ~talClustering and CoJlJiected SpeechRecognition, Proc. Sixth InternationalCongress on Acoustics, Vol. 2, pp. C-57 toC-60, Tokyo, 1968.
63
152. Reddy, D. Raj., Ann Robinson,Phoneme-to-Grapheme TraJlslatioJl ofEnglish, 1EEE Trans. Audio andElectyoacoustics, June 1968.
153. Reddy, D. Raj., Pierre Vicens,Procedure for SegJneJltatioJl ofCoJlJlected Speech, J. Audio Eng. Sot.,October 1968.
154. Roth, Bernard, Design, KiJleJnatics,aJid CoJitrol of Coinputer-coJitroIledManipulators, Proc. 6th All UnionConference on New Problems in Theory ofMachines 6? Mechanics, Leningrad, Jan.
_ 1971.
155. Sagan, Carl, J, Lederberg, E. Levinthal,L. Quam, R. Tucker, et al, VariableFeatures 011 Mars: PrelimiJJary MariJier9 Television Results, Icarus 17, pages346-372, 1972.
156. Samuel, Arthur, Studies in MachitleLearJliJlg UsiJlg the Game of Checkers,II-Recent Progress, IBM journal,November 1967.
157. Schank, Roger, Larry Tesler, ACoJlceptual Parser for NaturalLanguage, Proc. International JointConference on Artificial Intelligence,Washington, D.C., 1969.
158. Schank, Roger, Firidiug theConceptual Colltellt and IJiteJitiori illan Utterance in Natural LaJlguageConversation, Proc. 21JCAI, Brit. Comp.sot., 1971.
159. Schank, Roger, ConceptualDependency: a Theory of NaturalLaJiguage UJiderstaJidiJig, CognitivePsychology, Vol 3, No. 4, 1972.
160. Schank, Roger C., Neil M. Goldman,Theoretical Colisideratiorls itI TextProcessing, Conf. PYOC. Computer Text
64
Processing and Scientijc Research t/972),O.N.R., Pasadena, Calif., March 1973.
161. Schank, Roger C., Neil Goldman,Charles J. Rieger III, Chris Riesbeck,MARGIE: Memory, Atialysis, RespouseGeueratiou aud Iufereuce OII English,Proceedings of the Third InternationalJoint Conference on Artificial Intelligence,Stanford University, August 1973.
162. Schank, Roger C., Kenneth Colby (eds),Computer Models of Thought andLanguage, W. H. Freeman, SanFrancisco, 1973.
163. Schroll, G., A. Duffield, Carl Djerassi,Bruce Buchanan, G. Sutherland,Edward Feigenbaum, Joshua Lederberg,Applications of Artificial Intelligencefor Chemical Inference III. AliphaticEthers Diagnosed by Their LowResolutiou Mass Spectra aud NMRData, J. Amer. CAem. Sot., 91:26,
- December 1969.
164. Sheikh, Y., A. Buchs, A Delfino, BruceBuchanan, G. Sutherland, JoshuaLederberg, Applications of ArtificialIrrtelligence for Chemical Iuference V.AJI Approach to the ComputerGeueratiou of Cyclic Structures.Differentiatiou Between All the PossibleIsometric Ketones of CompositionC6H100, Organic Mass Spectrometry,Vol. 4 pp.493-501, 1970.
165. Silvestri, Anthony, Joseph Goodman,Digital Reconstructiou of HolographicImages, 1968 NEREM Record, IEEE,Vol. 10, pp. 118-119. 1968.
166. Slagle, James, Carl Farrell,Experitneuts in Automatic Learuiugfor a Multiputpose Heuristic Program,Comm. ACM, February 1971.
EXTERNAL PUBLICATIONS
167. Smith, D. H., B. G. Buchanan, R. S.Engelmore, A. M. Duffield, A. Yeo, E. A.Feigenbaum, J. Lederberg, C. D jerassi,Applicatioris of Artificial Intelligencefor Chemical Inference VIII. AJIapproach to the ComputerIuterpretatiorr of the High ResolutionMass Spectra of Complex Molecules.Structure Elucidation of EstrogenicSteroids, Journal of the AmericanChemical Society, 94, 5962-597 1, 1972.
168. Smith, David Canfield, Horace J. Enea,Backtracking in MLISPB, Proceedings of
_ the Third International joint Conferenceon Artificial Intelligence, StanfordUniversity, August 1973.
169. Smith, Leland, SCORE -- A Musician’sApproach to Computer Music, J. AudioEng. Sot., Jan./Feb. 1972.
170. Smith, Leland, Editing aud PrintingMusic by Computer, J. Music Theory,Fall 1973.
171. Sobel, Irwin, On CalibratingComputer Controlled Cameras forPerceiving 3-D Scenes, PYOC. Third ht.Joint Con/. on Artijictal Intelligence,Stanford U., 1973.
172. Sridharan, N., Search Strategies forthe Task of Organic ChemicalSynthesis, Proceedings of the ThirdInternational Joint Conference onArtificial Intelligence, StanfordUniversity, August 1973.
173. Sullivan, S. Brodsky and J., W-BOSOJICorltributiorl to the AuoinalousMagnetic Moinelit of the MUOJI, PhysRev 156, 1644, 1967.
174. Sutherland, Georgia, G. W. Evans, G.F.W allace, Simulation V sing DigitalComputers, Prentice-Hall, EngelwoodCliffs, N. J., 1967.
i
EXTERNAL PUBLICATIONS
175. Tenenbaum, Jay, et al, A Laboratoryfor Haud-eye Research, Pm. /F/PCongress, 197 1.
176. Tesler, Larry, Horace Enea, KennethColby, A Directed Graph Represerrtationfor Computer Simulation of BeliefSystems, Math. Bio. 2, 1968.
177. Tesler, Lawrence G., Horace J. Enea,David C. Smith, The LISPi’ PatternMatching System, Proceedings of theThird international Joint Conference onArtificial Intelligence, StanfordUniversity, August 1973.
1%. Waterman, Donald, GeneralizationLearning Techniques for Automatiugthe Learuiug of Heuristics, J. Artificialintelligence, Vol. 1, No. l/2.
179. Weyhrauch, Richard, Robin Milner,Program Sernautics aud Correctuess ina Mechanized Logic, Proc. USA-JapanComputer Conference, Tokyo, 1972.
180. Wilkes, Yorick, SemauticConsiderations irr Text Processing,Conj. Proc. Computer Text Processing andScienti,fic Research (J!72), O.N.R.,Pasadena, Calif., March 1973.
181. Wilks, Yorick, The Stauford MachineTranslatiou and Understanding Project,in Rustin (ed.) Natural LanguageProcessing, New York, 1973.
182. Wilks, Yorick, Uuderstaudiug WithoutProofs, Proceedings of the Thirdinternational Joint Conference onArti@iaI Intelligence, StanfordUniversity, August 1973.
183. Wilks, Yorick, Annette Herskovits, AnIntelligent Analyser and Generator ofNatural Language, Proc. ht. Conf. onComputational Linguistics, Pisa, Italy,Proceedings of the Third International
65
Joint Conference on Arti$cial intelligence,Stanford University, August 1973.
184. Wilks, Yorick, AJI ArtificialIntelligence Approach to MachineTranslation, in Schank and Colby (eds.),Computer Models of Thought andLanguage, W. H. Freeman, SanFrancisco, 1973.
185. Yakimovsky, Yoram, Jerome A.Feldman, A Semantics-Based DecisionTheoretic Region Analyzer, Proceedingsof the Third International JointConference on ArtiIciai Intelligence,
_ Stanford University, August 1973.
Appendix E
A. I. MEMO ABSTRACTS
In the listing below, there are up to threenumbers given for each Artificial IntelligenceMemo: an “AIM” number on the left, a “CS”(Computer Science) number in the middle,and a NTIS stock number (often beginning“AD...“) on the right. If there is no %” to theleft of the AIM number, then it is in stock atStanford at this writing and may berequested from:
Documentation ServicesArtificial Intelligence LaboratoryStanford UniversityStanford, California 94305
Alternatively, if there is an NTIS numbergiven, then the report may be ordered usingthat number from:
National Technical Information ServiceP. 0. Box 1553Springfield, Virginia 22 15 1
If there is no NTIS number given then theymay or may not have the report. Inrequesting copies in this case, give them boththe “AIM-” and “CS-nnn” numbers, with thelatter enlarged into the form “STAN-CS-yy-nnn”, where “yy” is the last two digits of theyear of publication.
Memos that are also Ph.D. theses are somarked below and may be ordered from:
University MicrofilmP. 0. Box 1346Ann Arbor, Michigan 48106
For people with access to the ARPANetwork, the texts of some A. I. Memos arestored on l ine in the Stanford A. I.Laboratory disk file. These are designatedbelow by “Diskfile: <file name>” appearing inthe header. See Appendix A for directionson how to access such files.
67
AIM-1John McCarthy,Predicate Calculus with ‘Undefined’ as aTruth-value,5 pages, March 1963.
T h e u s e o f p r e d i c a t e c a l c u l u s i n t hemathematical theory of computation and theproblems involved in interpreting theirvalues.
AIM-2John McCarthy,Situations, Actions, and Causal Laws,11 pages, July 1963.
A fo rma l t heo ry i s g iven conce rn ingsituations, causality and the possibility andeffects of actions is given. The theory isintended to be used by the Advice Taker, acomputer program that is to decide what todo by reasoning. Some simple examples aregiven of descriptions of situations anddeduct ions that cer ta in g o a l s c a n b eachieved.
AIM-3Fred Safier,‘The Mikado’ an an Advice Taker Problem,4 pages, July 1963.
The situation of the Second Act of ‘TheMikado’ is analyzed from the point of viewof Advice Taker formalism. This indicatesdefects still present in the language.
({AIM-4Horace Enea,Clock Function for LISP 1.5,2 pages, August 1963.
This paper describes a clock function forLISP 1.5.
Interesting work is being done inprogramming computers to solve problemswhich require a high degree of intelligencein humans. However, certain elementaryverbal reaesoning processes so simple theycan be ca.rried out by any non-feeble-mindedhuman have yet to be simulated by machineprograms.
T h i s p a p e r will discuss programs tomanipulate in a suitable formal language(most likely a part of the predicate calculus)common instrumental statements. The basicprogram will draw immediate conclusionsfrom a list of premises. These conclusionswill be either declarative or imperativesentences. When an imperative sentence isdeduced the program takes a correspondingaction. These actions may include printingsentences, moving sentences on lists, andreinitiating the basic deduction process onthese lists.
Facilities will be provided for communicationwith humans in the system via * manual
68 A. I. MEMO ABSTRACTS
AIM-5Horace Enea, Dean Wooldridge,Algebraic Simplication,2 pages, August 1963.
Herein described are proposed and effectedchanges and additions to Steve Russell’sMark IV Simplify.
:::AIM-6Dean Wooldridge,Non-printing Compiler,2 pages, August 1963.
A short program which redefines parts of theLISP 1.5 compiler and suppresses compilerprint out (at user’s option) is described.
AIM-7John McCarthy,Programs With Cotn tnou Sense,7 pages, September 1963.
intervention and display devices connected tothe computer.
aAIM-John McCarthy,Storage Conventions in LISP 2,? pages, September 1963.
Storage conventions and a basic set offunctions for LISP 2 are proposed. Since thememo was written, a way of supplementingthe features of this system with the uniquestorage of list structure using a hash rule forcomputing the address in a separate freestorage area for lists has been found.
t:AIM -9C. M. Williams,Computing Estimates for the Number ofBisections of an NxN Checkerboard for NEven,9 pages, December 1963.
This memo gives empirical justification forthe assumption that the number of bisectionsof an NxN (N even) checkerboard isapproximately given by the b inomialcoefficient (A, A/2) where 2A is the length ofthe average bisecting cut.
AIM-10Stephan R. Russell,Improvements in LISP Debugging,3 pages, December 1963.
Experience with writing large LISPprogrrams and helping students learningLISP suggests that spectacular improvementscan be made in this area. Theseimprovements are partly an elimination ofsloppy coding in LISP 1.5, but mostly anelaboration of DEFINE, the push down listbacktrace, and the current tracing facility.Experience suggests that these improvementswould reduce the number of computer runsto debug a program a third to a half.
A. I. MEMO ABSTRACTS
AIM-I 1Dean W oold ridge, Jr.,An Algebraic Simplify Program in LISP,57 pages, December 1963.
A program which performs ‘obvious’ (non-controversial) simplifying transformations onalgebraic expressions (written in LISP prefixnotation) is described. Cancellation ofinverses and consolidation of sums andproducts are the basic accomplishments ofthe program; however, if the user desirers todo so, he may request the program toperform special tasks, such as collect commonfactors from the products in sums or expandproducts. Polynomials are handled byroutines which take advantage of the specialform by polynomials; in particular, division(not cancellation) is always done in terms ofpolynomials. The program (run on the IBM7090) is slightly faster than a human;however, the computer does not need tocheck its work by repeating the simplification.
- Although the program is usable -- no bugsare known to exist -- it is by no means afinished project. A rewriting of the simplifysystem is anticipated; this will eliminate muchof the exis t ing redundancy and otherinefficiency, as well as implement an identity-recognizing scheme.
AIM-12Gary Feldman,Documentatiorl of the MacMahon SquaresProblem,4 pages, January 1964.
An exposition of the MacMahon Squaresproblem together with some ‘theoretical’results on the nature of its solutions and ashort discussion of an ALGOL programwhich finds all solutions are containedherein.
-
AIM- 13
69
Dean E. Wooldridge,The New LISP System (LISP 1,55),4 pages, February 1964.
The new LISP system is described.Although differing only slightly it is thoughtto be an improvement on the old system.
AIM-14John McCarthy,Computer Control of a Machine forExploring Mars,6 pages, January 1964.
Landing a 5000 pound package on Mars thatwould spend a year looking for life andmaking other measurements h a s b e e nproposed. We believe that this machineshould be a stored program computer withsense and motor organs and that the machineshould be mobile. We discuss the followingpoints:
1. Advantages of a computer controlledsystem.
2. What the computer should be like.3. What we can feasible do given the
present state of work on artificialintelligence.
4. A plan for carrying out research incomputer controlled experiments thatwill make the Mars machine as effectiveas possible.
AIM- 15Mark Finkelstein, Fred Safier,Axiomatization and Implementation,6 pages, June 1964.
An example of a typical Advice-Takeraxiomatization of a situation is given, andthe situation is programmed in LISP as anindication of how the Advice-Taker could beexpected to react. The situation chosen isthe play of a hand of bridge.
70 A. I. MEMO ABSTRACTS
AIM- 16 t(AIM- 19John McCarthy, Jan Hext,A Tough twt for Proof Procedures, Program mitlg Latlguages and Translation,3 pages, July 1964. 14 pages, August 1964.
It is well known to be impossible to tile withdominoes a checkerboard with two oppositecorners deleted. This fact is readily stated inthe first order predicate calculus, but theusual proof which involves a parity andcounting argument does not readily translateinto predicate calculus. We conjecture thatthis problem will be very difficult forprogrammed proof procedures.
AIM-17John McCarthy,Formal Description of the Game of Patlg-Ke,2 pages, July 1964.
The game of Pang-Ke is formulated in afirst-order-logic in order to provide grist forthe Advice-Taker Mill. The memo does notexplain all the terms used.
AIM-18Jan Hext,An Expressiorr Itlput Routine for LISP,5 pages, July 1964.
The expression input routine is a LISPf u n c t i o n , Mathread ( ] w i t h a s s o c i a t e ddefinitions, which reads in expressions suchas (A+S-F(X,Y,Z)). Its result is an equivalentS-expression. The syntax of allowableexpressions is given, but (unlike ALGOL’s) itdoes not def ine the precedence of theoperators; nor does the program carry outany explicit syntax analysis. Instead theprogram parses the expression according to aset of numerical precedence values, andreports if it finds any symbol out of context.
A notation is suggested for defining thesyntax of a language in abstract form,specifying only its semantic constituents. Asimple language is presented in this formand its semantic definition given in terms ofthese constituents. M e t h o d s a r e t h e ndeveloped for translating this language, firstinto LISP code and from there to machinecode, and for proving that the translation iscorrect.
AIM-20D. Raj. Reddy,Source Language Optimization of For-loops,37 pages, August 1964.
Program execution time can be reduced, by aconsiderable amount, by optimizing the ‘For-loops’ of Algol programs. By judicious useof index registers and by evaluating all thesub-expressions whose values are not alteredwithin the ‘For-loop’, such optimization canbe achieved.
In this project we develop an algorithm tooptimize Algol programs in list-structureform and generate a new source languageprogram, which contains the ‘desired contentsin the index registers’ as a part of the For-clause of the For-statement and additionalstatements for evaluating the sameexpressions outside the ‘For-loop’ Thisoptimization is performed only for theinnermost ‘For-loops’.
The program is written entirely in LISP.Arrays may have any number of subscripts.Further array declarations may have variabledimensions. (Dynamic allocation of storage.)The program does not t ry to opt imizearithmetic expressions. (This has alreadybeen extensively investigated.)
A. I. MEMO ABSTRACTS
AIM-2 1R. W. Mitchell,LISP 2 Specifications Proposal,12 pages, August 1964.
Specifications for a LISP 2 system areproposed. The source language is basicallyAlgol 60 extended to include list processing,input/output and language extensionfacilities. The system would be implementedwith a source language translator andoptimizer, the output of which could beprocessed by either an interpreter or acompiler. The implementation is specifiedfor a single address computer with particularreference to an IBM 7090 where necessary .
Expected efficiency of the system for listprocessing is significantly greater than theLISP 1.5 compiler. For execution of numericalgorithms the systems should be becomparable to many current “algebraic”compilers. Some familiarity with LISP, 1.5Algol and the IBM 7090 is assumed.
AIM-22Richard Russell,Kalah -- the Game and the Program,13 pages, September 1964.
A description of Kalah and the Kalahprogram, including sub-routine descriptionsand operating instructions.
AIM-23Richard Russell,Improvements to the_Kalah Program12 pages, September 1964.
Recent improvements to the Kalah programare listed, and a proposal for speeding up theprogram by a factor of three is discussed.
AIM-24John McCarthy,A Formal Description of a Subset ofALGOL,43 pages, September 1964.
-
71
We describe Microalgol, a trivial subset ofAlgal, by means of an interpreter. Thenotions of abstract syntax and of ‘state of thecomputation’ permit a compact description ofboth syntax and semantics. We advocate anextension of this technique as a general wayof describing programming language.
AIM-25Richard Mansfield,A Formal System of Computation,7 pages, September 1964.
We discuss a tentative axiomatization for aformal system of computation and within thissystem we prove certain propositions aboutthe convergence of recursive definitionsproposed by J. McCarthy.
AIM-26D. Raj. Reddy,Expertllleuts 011 Automatic SpeechRecognition by a Digital Computer,19 pages, October 1964.
Speech sounds have in the past beeninvestigated with the aid of spectographs, vo-coders and other analog devices. With theavailability of digital computers withimproved i-o devices such as Cathode Raytubes and analog digital converters, it hasrecently become practicable to employ thispowerful tool in the analysis of speechsounds.
Some papers have appeared in the recentliterature reporting the use of computers inthe determination o f t h e f u n d a m e n t a lfrequency and for vowel recognition. Thispaper discusses the details and results of apreliminary investigation conduc ted a tStanford. It includes various aspects ofspeech sounds such as waveforms of vowelsand consonants; de terminat ion of afundamental of the wave; Fourier (spectral)analysis of the sound waves formatdetermination, simple vowel recognitionalgorithm and synthesis of sounds. All wereobtained by the use of a digital computer.
72 A. I. MEMO ABSTRACTS
AIM-27John McCarthy,A Proof -checker for Predicate Calculus,7 pages, March 1965.
A program that checks proofs in J. A.Robinson’s formulation of predicate calculushas been programmed in LISP 1.5. Theprogram is available in CTSS at ProjectMAC and is also available as a card deck.The program is used for class exercises atStanford.
AIM -28John McCarthy,Problems in the Theory of Computatiotl,7 pages, March 1965.
The purpose of this paper is to identify anddiscuss a number of theoretical problemswhose solutions seem feasible and likely toadvance the practical art of computation.The problems that will be discussed includethe following:
1. Semantics of programming languages.What do the strings of symbols representingcomputer programs, statements, declarations,labels, etc., denote? How can the semantics ofprogramming languages be describedformally?
2. Data spaces. What are the spaces of dataon which computer programs act and howare they built up up from simpler spaces?
3. How can time d e p e n d e n t ’ a n d -simultaneous processes be described?
4. Speed of computation. What can be saidabout how much computation is required tocarry out certain processes?
5. Storage of information. How caninformation be stored so that items identicalor similar to a given item can be retrieved?
appropriate domain for computationsdescribed by productions or other dataformat recogn izers?
7. What are the appropriate formalisms forwriting proofs that computer programs areequivalent?
8. In the view of Codel’s theorem that tellsus that any formal theory of computationmust be incomplete, what is a reasonableformal system that will enable us to provethat programs terminate in practical cases?
AIM-29Charles M. Willtams,Isolation of Important Features of aMultitoued Picture,9 pages, January 1965.
A roughly successful attempt is made toreduce a multi-toned picture to a two-toned.(line drawing) representation capable ofbeing recognized by a human being.
AIM-30Edward A. Feigenbaum, Richard W. Watson,AJI Initial Problem Statement for aMachine Induction Research Project,8 pages, April 1965.
A brief description is given of a researchproject presently getting under way. Thisproject will study induction by machine,using organic chemistry as a task area.Topics for graduate student research relatedto the problem are listed.
AIM-31John McCarthy,PIaIls for the Stanford Artificialllitelligeme Project,17 pages, April 1965.
The following is an excerpt from a proposalto ARPA and gives some of the project plansfor the near future.
6. Syntax directed computation. What is the
A. I. MEMO ABSTRACTS 73
AIM-32Harry Ratchford,The 138 Analog Digital Converter,9 pages, May 1965.
A discuss ion of the programming andhardware characteristics of the analog todigital converter on the PDP-1 is given;several sample programs are also presented,
rr AIM-33
Barbara Huberman,The Advice Taker and GPS,8 pages, June 1965.
Using the formalism of the Newell-Shaw-Simon General Problem Soiver to solveproblems expressed in McCarthy’s AdviceTaker formal ism is d iscussed. Somerevisions of the formalism of can and causedescribed in AI Memo 2 are proposed.
AIM-34Peter Cara h,4 Television Camera Interface for thePDP-I,8 pages, June 1965.
This paper is a discussion of several methodsfor the connection of a television camera tothe PDP- 1 computer. Three of these methdsare discussed in detail and have in commontha.t only a 36 bit portion of any horizontalscanning line may be read and thisinformation is read directly into the workingregisters of the computer. The fourthinvolves a data channel to read informationdirectly into the core memory of thecomputer, and is mentioned only in passing.The major concepts and some of the detailsof these methods are due to Marvin Minsky.
::tAIM-3_lFred Safier,Simple Simon,17 pages, June 1965.
SIMPLE SIMON is a program which solves
I -
the problem of finding an object satisfying apredicate from a list of facts. It operates bybackward chaining. The rules of procedureand heuristics are discussed and the structureof the program is outlined.
AIM-36James Painter,Utilization of a TV Camera on the PDP-1,6 pages, September 1965.
A description of the programming requiredto utilize the TV camera connected to thePDP-1 and of the initial collection ofprograms.
((AIM-37Knut Korsvold,An 011, Lisle Algebraic SimplificatiorlProgram,36 pages, November 1965.
We describe an on-line program for algebraicsimplification. The program is written inLISP 1.5 for the Q32 computer at SystemDevelopment Corporation in Santa Monica,California. The program has in its entiretybeen written and debugged from a teletypestation at Stanford University.
AIM-38Donald A. Waterman,A Filter for a Machine Induction System,19 pages, January 1966.
This report contains current ideas about theMachine Induction Research Project, andattempts to more clearly define some of theproblems involved. In particular, the on-linedata acquisition problem, the filter, and theinductive inference problem associated withthe filter are discussed in detail.
AIM-39Karl Pingle,A Program to Find Objects in a Picture,22 pages, January 1966.
74 A. I. MEMO ABSTRACTS
A program is described which traces aroundobjects in a picture, using the picture scannerattached to the PDP-1 computer, and fitscurves to the edges.
AIM-40 CS-38 AD662880John McCarthy, James Painter,Correctness of a Compiler for ArithmeticExpressions,13 pages, April 1966.
This is a preprint of a paper given at theSymposium of Mathematical Aspects ofComputer Sc ience o f the AmericanMathematical Society held April 7 and 8,1966. It contains a proof of the correctnessof a compiler for arithmetic expressions.
:::AIM-4 1Phil Abrams, Dianna Rode,A Proposal for a Proof-checker for CertaihjAxiomatic Systems,10 pages, May 1966.
A proposed design for a proof-checker tooperate on many axiomatic domains ispresented. Included are descriptions of theorganization and operation of the program tobe written for the PDP-6.
AIM-42Karl Pingle,A Proposal for a Visual Input Routine,11 pages, June 1966.
Some comments are made o n thecharacteristics believed desirable in the nexteye for the Stanford Artificial IntelligenceProject and a proposal is given for aprogram to input scenes using the eye.
:::AIM-43 ’ cs-49D. Raj.-Reddy,
SS640-836
An Approach to Computer SpeechRecognition by Direct Analysis of theSpeech Wave,Thesis: Ph.D. in Computer Science,144 pages, September 1966.
A s y s t e m f o r o b t a i n i n g a p h o n e m i ctranscription from a connected speech sampleentered into the computer by a microphoneand an analog-to-digital converter isdescribed. A feature-extraction programdivides the speech utterance into segmentsapproximately corresponding to phonemes,determine pitch periods of those segmentswhere pitch analysis is appropriate, andcomputes a list of parameters for eachsegment. A classification program assigns aphoneme-group label (vowel-like segment,fricative-like segment, etc.) to each segment,determines whether a segment should beclass i f ied as a phoneme or whether i trepresents a phoneme boundary between twophonemes, and then assigns a phoneme labelto each segment that is not rejected as beinga phoneme boundary. About 30 utterancesof one to two seconds duration were analyzedusing the above programs on a ninterconnnected I B M 7090-PDP- 1 s y s t e m . ’Correct identification of many vowel andconsonantal phonemes was achieved for asingle speaker. The time for analysis of eachutterance was about 40 times real time. Theresults were encouraging and point to a newdirection in speech research.
({AIM-44James Painter,Semantic Correctness of a Compiler for anAlgol-like Language,Thesis: Ph.D. in Computer Science,130 pages, revised March 1967.
This is a semantic proof of the correctness ofa compiler. T h e abs t rac t syn tax andsemantic definition are given for thelanguage Mickey, an extension of Micro-algol. The abstract syntax and semantics aregiven for a hypothetical one-register single-address computer with 14 operations. Acompiler, using recursive descent, is defined.Formal definitions are also given for statevector, a and c functions, and correctness of acompiler. Using these definitions, thecompiler is proven correct.
A. I. MEMO ABSTRACTS 75
AIM-45Donald Kaplan,Some Completeness Results in theMathematical Theory of Computation,22 pages, October 1966.
A formal theory is described whichincorporates the ‘assignment’ function a(i, k,psi) and the ‘contents’ function c(i, psi). Theaxioms of the theory are shown to comprise acomplete and consistent set.
notably by N. R. Hanson, and S. E.Toulmin. While a logic of discovery isgenerally understood to be an algorithm forformulating hypotheses, other concepts havebeen suggested. Chapters V and VI exploretwo of these: (A) a set of criteria by which ahypothesis could be judged reasonable, and(B) a set of rational (but not necessarilyeffective) methods for formulatinghypotheses.
::!AIM-46 cs-50Staffan Persson,
PB 176761AIM-48Donald M. Kaplan,
Some Sequence Extrapolating Programs: aStudy of Representation and Modeling inInquirilig Systems,Thesis: Ph.D. in Computer Science,176 pages, September 1966.
CorrectlIess of a Compiler for Algal-likePrograms,46 pages, July 1967.
The purpose of this thesis is to investigatethe feasibility of designing mechanizedinquiring-systems for finding suitablerepresentations of problems, i.e., to performthe ‘creative’ task of finding analogies.Because at present a general solution to thisproblem does not seem to be within reach,the feasibility of mechanizing a particularrepresentational inquirer is chosen as areasonable first step towards an increasedunderstanding of the general problem. It isindicated that by actually designing,programming and running a representationalinquirer as a program for a digital computer,a severe test of its consistency and potentialfor future extensions can be performed.
A compiling algorithm is given which mapsa class of Algol-like programs into a class ofmachine language programs. The semantics,i. e., the effect of execution, of each class isspecified, and recursion induction used toprove that program semantics is preservedunder the mapping defined by the compilingalgorithm.
AIM-49Georgia Sutherland,DENDRAL -- a Computer Program forGenerating aud Filtering ChemicalStructures,34 pages, February 1967.
+AIM-47Bruce Buchanan,Logics of Scientific Discovery,Thesis: Ph.D. in PAiiosophy UC. Berkeley,2 10 pages, December 1966.
The concept of a logic of discovery isdiscussed from a philosophical point of view.Early chapters discuss the concept ofdiscovery itself, some arguments have beenadvanced against the logics of discovery,
A computer program has been written whichcan generate all the structural isomers of achemical composition. The generatedstructures are inspected for forbiddensubstructures in order to eliminate structureswhich are chemically impossible from theoutput. In addition, the program containsheuristics for determining the most plausiblestructures, for utilizing supplementary data,and for interrogating the on-line user as todesired options and procedures. Theprogram incorporates a memory so that pastexperiences are utilized in later work.
-
76 A. I. MEMO ABSTRACTS
AIM-50Anthony C. Hearn,Reduce Users’ Manual,53 pages, February 1967.
REDUCE is a program designed for generalalgebraic computations of interest tophysicists and engineers. Its capabilitiesinclude:
1) expansion and ordering of rationalfunctions of polynomials,
2) symbolic differentiation,3) substitutions in a wide variety of forms,4) reduction of quotients of polynomials by
cancellation of common factors,5) calculation of symbolic determinants,6) calculations of interest to high energy
physicists including spin l/2 and spin 1algebra.
The program is written completely in thelanguage LISP 1.5 and may therefore be runwith little modification on any computerpossessing a LISP 1.5 compiler or interpreter.
AIM-51Lester D. Earnest,Choosing an eye for a Coinpute’r,I54 pages, April 1967.
In order for a computer to operate efficientlyin an unstructured environment, it must haveone or more manipulators (e. g., arms andhands) and a spatial sensor analogous to thehuman eye. Alternative sensor systems arecompared here in their performance oncertain simple tasks. Techniques fordetermining color, texture, and depth ofsurface elements are examined. Sensingelements considered include thep hotomultiplier, image dissector, imageorthicon, vidicon, and SEC camera tube.Performance measures strongly favor a new(and undemonstrated) configuration that maybe termed a laser jumping spot system.
AIM-52Arthur L. Samuel,Some Studies in Machine Learning Usirlgthe Came of Checkers II - Recent Progress,48 pages, June 1967.
A new signature table technique is describedtogether with an improved book learningprocedure which is thought to be muchsuperior to the linear polynomial methoddescribed earlier. Full use is made of the socalled alpha-beta pruning and several formsof forward pruning to restrict the spread ofthe move tree and to permit the program tolook ahead to a much greater depth than itotherwise could do. While still unable tooutplay checker masters, the programs’splaying ability ,has been greatly improved.Some ,of these newer techniques should beapplicable to problems of economicimportance.
AIM-53William Weiher,The PDP-6 Proof Checker,47 pages, June 1967.
A descrription is given for the use of a proofchecker for propositional calculus. Anexample of its use as well as the M and Sexpressions for the proof checker are alsoincluded.
AIM-54Joshua Lederberg, Ed ward A. Feigen baum,Mechanization of Inductive Inference inOrganic Chemistry,29 pages, August 1967.
A computer progra.m for formulatinghypotheses in the area of organic chemistryis described from two standpoints: artificialintelligence and organic chemistry. TheDendral Algorithm for uniquely representingand ordering chemical structures defines thehypothesis-space; but heuristic searchthrough the space is necessary because of itssize. Both the algorithm and the heuristics
A. I. MEMO ABSTRACTS
are described explicitly but without referenceto the LISP code in which these mechanismsare programmed. Within the program someuse has been made of man-machineinteraction, pattern recognition, learning, andtree-pruning heuristics as well as chemicalheuristics which allow the program to focusits attention on a subproblem to rank thehypotheses in order of plausibility. Thecurrent performance of the program isillustrated with selected examples of actualoutput showing both its algorithmic andheuristic aspects. In addition some of themore important planned modifications arediscussed.
AIM-55Jerome Feldman,First Thoughts of Grammatical Inference,18 pages, August 1967.
A number of issues relating to the problemof inferring a grammar are disscussed. Astrategy for grammatical inference ispresented and its weaknesses and possibleimprovements are discussed. This is aworking paper and should not bereproduced, quoted or believed without theauthor’s permission. .
AIM-56William Wichman,Use of Optical Feedback in the ComputerControl of an Arm,Thesis: Eng. in Eiectrical Engineering,69 pages, August 1967.
This paper reports an experimentalinvestigation of the application of visualfeedback to a simple computer-controllerblock-stacking task. The system uses avidicon camera to examine a table containingtwo cubical blocks, generating a datastructure which is analyzed to determine theposition of one block. An electric arm picksup the block and removes it from the scene,then after the program locates the secondblock, places the first on top of the second.
77
Finally, the alignment of the stack isimproved by analysis of the relative positionerror as seen by the camera. Positions aredetermined throughout by perspectivetransformation of edges detected from asingle viewpoint, using a support hypothesisto supply sufficient information on depth.The Appendices document a portion of thehardware used in the project.
AIM-57Anthony C. Hearn,REDUCE, a User-oriented InteractiveSystetn for Algebraic Slmpliflcatlon,69 pages, October 1967.
This paper describes in outline the structureand use of REDUCE, a program designedfor large-scale algebraic computations ofinterest to applied mathematicians, physicists,and engineers. The capabilities of the systeminclude:
1) expansion, ordering and reduction ofrational functions of polynomials,
2) symbol differentiation,3) substitutions for variables and
expressions appearing in otherexpressions,
4) simplification of symbolic determinantsand matrix expressions,
5) tensor and non-commutative algebraiccalculations of interest to high energy ,physicists.
In addition to the operations of addition,subtraction, multiplication, division,numerical exponentiation numericalexponentiation and differentiation, it ispossible for the user to add new operatorsand define rules for their simplification.Derivations of these operators may also mayalso be defined.
The program is written complete in thelanguage of LISP 1.5 and is organized so asto minimize the effort required intransferring from one LISP system toanother.
78 A. I. MEMO ABSTRACTS
Some particular problems which have arisenin using REDUCE in a time-sharingenvironment are also discussed.
AIM-58Monte D. Callero,An Adaptive Command arld ControlSystem Utilizing Heuristic LearningProcesses,Thesis: Ph.D. in Operations Research,161 pages, December 1967.
procedure with the result that the decisionfunction coefficients converged under thelearning process and the decision processbecome increasingly effective.
::aAIM-59Donald M. Kaplan,A Formal Theory Conceruillg theEquivalence of Algorithms,20 pages, May 1968.
The objectives of the research reported hereare to develop an automated decision processfor real time allocation of defense missiles toattacking ballistic missiles in general war andto demonstrate the effectiveness of applyingheuristic learning to seek optimality in theprocess. The approach is to model andsimulate a missile defense environment andgenerate a decision procedure featuring aself-modifying, heuristic decision functionwhich improves its performance withexperience. The goal of the decision processtha.t chooses between the feasible allocationsis to minimize the total effect of the attack,measured in cumulative loss of target value.The goal is pursued indirectly by consideringthe more general problem of maintaining astrong defense posture, the ability of thedefense system to protect the targets fromboth current and future loss.
Using simulation and analysis, a set ofcalculable features are determined whicheffectively reflect the marginal deteriorationof defense posture for each allocation in atime interval. A decision function, a linearpolynomial of the features, is evaluated foreach feasible allocation and the allocationhaving the smallest value is selected. Aheuristic learning process is incorporated inthe model to evaluate the performancee ofthe decision process and adjust the decisionfunction coefficients to encourage correctcomparison of alternative allocations.Simulated attacks presenting typical defensesituations were cycled against the decision
Axioms and rules of inference are given forthe derivation of equivalence for algorithms.The theory is shown to be complete forcertain subclasses of algorithms, an,d severalapplications of the theory are illustrated.This paper was originally presented at theMathematical Theory of ComputationConference, IBM Yorktown Heights,November 27-30, 1967.
AIM-60 cs-101Donald M. Kaplan,
AD672923
The Formal Theoretic Allalysis of StrongEquivalence for Elernelltal Programs,Thesis: Ph.D. in Computer Science,263 pages, June 1968.
The syntax and semantics is given forelemental programs, and the strongequivalence of these simple ALGOL-likeflowcharts is shown to be undecidable. Aformal theory is introduced for derivingstatements of strong equivalence, and thecompleteness of this theory is obtained forvarious sub-cases. Several applications ofthe theory are discussed. Using a regularexpression representation for elementalprograms and an unothodox semantics forthese expressions, several strong equivalencedetecting procedures are developed. Thiswork was completed in essentially its presentform March, 1968.
A. 1. MEMO ABSTRACTS 79
::eAIM-6 1Ta kayasu Ito,Notes on Theory of Computation andPattern Recognition,144 pages, May 1968.
This is a collection of some of the author’sraw working notes during the periodDecember 1965 - October 1967 besides theintroduction. They have been privately orinternally distributed for some time. Portionsof this work have been accepted forpublication; others are being developed forsubmission to journals. Some aspects andideas have been referred to and used,sometimes without explicit references, andothers are developed by other researchersand the author. Hence we have decided topublish this material as a Computer ScienceTechnical Report, although the author isplanning to submit all of these works to somejournals, adding several new results (notmentioned in this report), improvingnotations, definitions and style of presentationin -some parts and reformulating completelyin other parts.
AIM-62Bruce Buchanan, Georgia Sutherland,Heuristic Dendral: a Program forGeueratittg Explanatory Hypotheses inOrganic Chemistry,76 pages, July 1968.
A computer program has been written whichcan formulate hypotheses from a given set ofscientific data. The data consist of the massspectrum and the empirical formula of anorganic chemical compound. The hypotheseswhich were produced describe molecularstructures which are plausible explanationsof the data. The hypotheses are generatedsystematically within the program’s theory ofchemical stability and within limitingconstraints which are inferred from the databy heuristic rules. The program excludeshypotheses inconsistent with the data andlists its candidate explanatory hypotheses in
order of decreasing plausibility. Thecomputer program is heuristic in that itsearches for plausible hypotheses in a smallsubset of the total hypothesis space accordingto heuristic rules learned from chemists.
AIM-63Donald M. Kaplan,Regular Expressions and the Equivalenceof Programs,42 pages, July 1968.
The strong equivalence of ALGOL-likeprograms is, in general, an undecidableproperty. Several mechanical procedures arediscussed which nevertheless are useful inthe detection of strong equivalence, Thesemethods depend on a regular expressionrepresentation of programs. An unorthodoxsemantics for these expressions is introducedwhich appreciably adds to the ability todetect strong equivalence. Several othermethods of extending this ability are alsodiscussed.
({AIM-64Zohar Manna,Formalization of Properties of Programs,18 pages, July 1968.
Given a program, an algorithm will bedescribed for constructing an expression,such that the program is valid (i.e.,terminates and yields the right answer) ifand only if the expression is inconsistent.Similar result for the equivalence problem ofprograms is given. These results suggest anew approach for proving the validity andequivalence of programs.
AIM-65 CS- 106 AD67397 1Barbara J. Huberman,A Program to Play Chess end Games,Thesis: Ph.D. in Computer Science,168 pages, August 1968.
A program to play chess end games isdescribed. The model used in the program is
80 A. I. MEMO ABST R AC T S
very close to the model assumed in chessbooks. Embedded in the model are twopredicates, BETTER and WORSE, whichcontain the heuristics of play, different foreach end game. The definitions of BETTERand WORSE were obtained by programmertranslation from the chess books.
The program model is shown to be a goodone for chess and games by the successachieved for three end games. Also themodel enables us to prove that the programcan reach checkmate from any startingposition. Insights about translation frombook problem solving methods into computerprogram heuristics are discussed; they areobtained by comparing the chess bookmethods with definitions of BETTER andWORSE, and by considering the difficultyencountered by the programmer when doingthe translation.
?-::AIM -66Jerome A. Feldman, Paul D. Rovner,AII Algol-based Associative Language,31 pages, August 1968.
A high-level programming language for largecomplex relational structures has beendesigned and implemented. The underlyingrelation al data. structure has beenimplemented using a hash-coding technique.The discussion includes a comparison withother work and examples of a.pplications ofthe language. A version of this paper willappear in the Communications of the ACM.
:::AIM-67 AD680487Edward A. Feigenbaum,Artificial Intelligence: Themes in theSecond Decade,39 pages, August 1968.
In this survey of Artificial Intelligenceresearch, the substantive focus is heuristicprogramming, problem solving, and closelyassociated learning models. The focus intime is the period 1963-1968. Brief tours are
made over a variety of topics: generality,integrated robots, game playing, theoremproving, semantic information processing, etc.
One program, which employs the heuristicsearch paradigm to generate explanatoryhypotheses in the analysis of mass spectra oforganic molecules, is described in some detail.The problem of representation for problemsolving systems is discussed. Various centersof excellence in the Artificial Intelligenceresearch area are mentioned. A bibliographyof 76 references is given.
AIM-68Zohar Manna, Amir Pnueli,The Validity Problem of the 91-function,20 pages, August 1968.
Several methods for proving the weak andstrong validity of algorithms are presented.
For proving the weak validity (i.e.,correctness) we use satisfiability methods,while proving the strong validity (i.e.,termination and correctness) we useunsatisfiability methods.
Two types of algorithms are discussed:recursively defined functions and programs.
Among the methods we include knownmethods due to Floyd, Manna, andMcCarthy. All the methods will beintroduced quite informally by means of anexample (the 9Lfunction).
t<AIM-69 AD677588John McCarthy, Edward Feigenbaum,Arthur Samuel,Project Technical Report,90 pages, September 1968.
Recent work of Stanford ArtificialIntelligence Project is summarized in severalareas:Scientific Hypothesis FormationSymbolic ComputationHand-Eye Systems
A. I. MEMO ABSTRACTS 81
Computer Recognition of SpeechBoard G amcsOther Projects
Hz, 900-2200 Hz, 2200-5000 Hz). Details ofthe design and implementation o f t hehardware device are given.
AIM-70Anthony C. Hcarn,The Problem of Substitution,14 pages, December 1968.
AD680072 (<AIM-72 CS- 116Donald L. Pieper,
AD680036
One of the most significant features ofprograms designed for non-numericcalculation is that the size of expressionsmanipulated, and hence the amount ofstorage necessary, changes continually duringthe exekution of the program. It is, therefore,usually not possible for the user to knowahead of time whether the calculation will infact fail because of lack of availablecomputer memory. The key to keeping boththe size of intermediate expressions andoutput under control often lies in the mannerin which substitutions for variables andexpressions declared by the programmer areimplemented by the system. In this papervarious methods which have been developedto perform these substitutions in the author’sown system REDUCE are discussed. A briefdescription of the REDUCE system is alsogiven.
The Kinematics of Manipulators underComputer Control,Thesis: Ph.D. in Mechanical Engineering,157 pages, October 1968.
The kinematics of manipulators are studied.A model is presented which allows for thesystematic description of new and existingmanipulators.
Six degree-of-freedom manipulators arestudied. Several solutions to the problem offinding the manipulator configuration leadingto a specified position and orientation arepresented. Numerical as well as explicitsolutions are given. The problem ofpositioning a multi-link digital arm is alsodiscussed.
AIM-7 1 AD677520Pierre V icens,Preprocessing for Speech Analysis,33 pages, October 1968.
Given the solution to the position problem,as a see of heuristics is developed for movinga six degree-of-freedom manipulator from aninitial position to a final position through aspace containing obstacles. This results in acomputer program shown to be able to directa manipulator around obstacles.
This paper describes a procedure, and itshardware implementation, for the extractionof significant parameters of speech. Theprocess involves division of the speechspectrum into convenient frequency bands,and calculation of amplitude and tero-crossing parameters in each of these bandsevery 10 ms. In the software implementation,a smooth function divides the speechspectrum into two frequency bands (aboveand below 1000 Hz). In the hardwareimplementation, the spectrum is divided intothree bands using bandpass filters (150-900
+AIM-73 AD678878John McCarthy, Patrick Hayes,Some Philosophical Problems From theStandpoint of Artificial Intelligence,51 pages, November 1968.
A computer program capable of actingintelligently in the world must have a generalrepresentation of the world in terms of whichits inputs are interpreted. Designing such aprogram requires commitments about whatknowledge is and how it is obtained. Thussome of the major traditional problems ofphilosophy arise in artificial intelligence.
82 A. I. MEMO ABSTRACTS
More specifically, we want a computerprogram that decides what to do by inferringin a formal language that a certain strategywill achieve its assigned goal. This requiresformalizing concepts of causality, ability, andknowledge. Such formalisms are alsoconsidered in philosophical logic.
The first part of the paper begins with aphilosophical point of view that seems toarise naturally once we take seriously theidea of actually making an intelligentmachine. We go on to the notions ofmetaphysically and epistemologicallyadequate representations of the world andthen to an explanation of can, causes, andknows, in terms of a representation of theworld by a system of interacting automata. Aproposed resolution of the problem offreewill in a deterministic universe and ofcounterfactual conditional sentences ispresented.
The second part is mainly concerned withformalisms within which it can be provedtha.t a strategy will achieve a goal. Conceptsof situation, fluent, future operator, action,strategy, result of a strategy and knowledgeare formalized. A method is given ofconstructing a sentence of first order logicwhich will be true in all models of certainaxioms if and only is a certain strategy willachieve a certain goal.
The formalism of this paper represents anadvance over (McCarthy 1963) and (Green1968) in that it permits proof bf the -correctness of strategies that contain loopsand strategies that involve the acquisition ofknowledge, and it is also somewhat moreconcise.
The third part discusses open problems inextending the formalism of Part II.
The fourth part is a review of work inphilosophical logic in relation to problems ofArticial Intelligence and discussion of
previous efforts to program generalintelligence from the point of view of thispaper. This paper is based on a talk givento the 4th Machine Intelligence Workshopheld at Edinburgh, August 12-21, 1968, andis a preprint of a paper to be published inMachine Intell igence 4 ( E d i n b u r g hUniversity Press, 1969).
gaAIM-74 CS-118 AD68 1027Donald Waterman,Machine Learning of Heuristics,Thesis: Ph.D. in Computer Science,? pages, December 1968.
The research reported here is concerned withdevising machine-learning techniques whichcan be applied to the problem of automatingthe learning heuristics.
<{AIM-75Roger C. Schank,A Notion of Linguistic Concept: a Preludeto Mechanical Translation,2 1 pages, December 1968.
The conceptual dependency framework hasbeen used as an automatic parser for naturallanguage. Since the parser gives as output aconceptual network capable of expressingmeaning in language-free terms, it is possibleto regard this as an interlingua. If aninterlingua is actually available how mightthis interlingua be used in translation? Theprimary problem that one encounters is thedefinition of just what these concepts in thenetwork are. A concept is defined as anabstraction in terms of percepts and thefrequency of connection of other concepts.This definition is used to facilitate theunderstanding of some of the problems inparaphrasing and translation. Themotivation for this abstract definition oflinguistic concept is discussed in the contextof its proposed use.
::tAIM-76Roger C. Schank,A Conceptual Parser for Natural Language,22 pages, December 1968.
c
This paper describes an operable automaticparser for natural language. The parser isnot concerned with producing the syntacticstructure of an input sentence. Instead, it is aconceptual parser, concerned withdetermining the underlying meaning of theinput. The output of the parser is a networkof concepts explicating the conceptualrelationships in a piece of discourse. Thestructure of this network is language-free;thus, sentences in different languages orparaphrases within the same language willparse into the same network. The theorybehind this representation is outlined in thispaper and the parsing algorithm is explainedin some detail.
DESCRIPTORS: computationalconcepts, linguistic research,understanding.
linguistics,computer
AIM-77Joseph D. Becker,The Modelirrg of Simple Aualogic andInductive Processes irr a Setnarrtic MemorySystem,21 pages, January 1969.
In this paper we present a general datastructure for a semantic memory, which isdistinguished in that a notion of consequence(temporal, ““causal, logical, or behavioral,depending on interpretation) is a primitive ofthe data representation. The same item of adata may at one time serve as a logicalimplication, and at another time as a‘pattern/action’ rule for behavior.
We give a definition of ‘analogy’ between Machine perception of vision and speech,items of semantic information. Using the thus, provides a problem domain for testingnotions of consequence and analogy, we the adequacy of the models of representationconstruct an inductive process in which (McCarthy and Hayes), planning heuristicgeneral laws are formulated and verified on selection (Minsky, Newell and Simon), and
83
the basis of observations of individual cases.We illustrate in detail the attainment of therule ‘Firemen wear red suspenders’ by thisprocess.
Finally, we discuss the relationship betweenanalogy and induction, and their use inmodeling aspects of ‘perception’ and‘understanding’.
AIM-78D. Raj. Reddy,011 the use of Environl.nental, Syntacticand Probalistic Constraints in Visioll alldSpeech,23 pages, January 1969.
In this paper we consider both vision andspeech in the hope that a unified treatment,illustrating the similarities, would lead to abetter appreciation of the problems, andpossibly programs which use the samesuperstructure. We postulate a generalperceptual system and illustrate how variousexisting systems either avoid or ignore someof the difficult problems that must beconsidered by a general perceptual system.The purpose of this paper is to point outsome of the unsolved problems, and tosuggest some heuristics that reflectenvironmental, syntactic, and probabilisticconstraints useful in visual and speechperception by machine. To make effectiveuse of these heuristics, a program mustprovide for
1. An external representation of heuristicsfor ease of man-machine communication
2. An internal representation of heuristicsfor effective use by machine
3. A mechanism for the selection ofappropriate heuristics for use in a givensituation.
84 A. I. MEMO ABSTRACTS
generalization learning (Samuel); a domain inwhich (perceptual) tasks are performed bypeople easily and without effort.
AIM-79 AD6856 11D. Raj. Reddy, Richard B. Neely,Cortex tual Analysis of Phonemes ofEnglish,71 pages, January 1969.
It is now well known that the acousticcharacteristics of a Phoneme depend on boththe preceding and following phonemes. Thispaper ,provides some needed contextual andprobabilistic da.ta about trigram phonemicsequences of spoken English. Since there areapproximately 40t3 such sequences, one mustdiscover a.nd study only the more commonlyoccurring sequences. To this purpose, threetypes of tables are presented, viz.,
a. Commonly occurring trigram sequencesof the form /abc/ for every phoneme/b/.
b. Commonly occurring sequences /abc/ forevery pair of phonemes /a/ and /cl.
c. Commonly occurring word boundarysequences of the form /-abl and lab-/where /-/ represents the silencephoneme.
Entries of the above tables contain examplesof usage and probabilities of occurrence foreach such sequence.
:::AIM -80 AD6856 12Georgia Sutherland,Heuristic Deudral: a-Family of LISPPrograms,46 pages, March 1969.
The Heuristic Dendral program for Machines that may be said to functiongenerating explanatory hypotheses in organic intelligently must be able to understandchemistry is described as an application of questions posed in natural language. Sincethe programming language LISP. The natural language may be assumed to have andescription emphasizes the non-chemical underlying conceptual structure, it isaspects of the program, particularly the desirable to have the machine structure its‘topologist’ which generates all tree graphs of own experience, both linguistic anda collection of nodes. nonlinguistic, in a manner concomitant with
$AIM-8 1 AD6856 13David Luckham,Refinement Theorems in ResolutionTheory,3 1 pages, March 1969.
The paper discusses some basic refinementsof the Resolution Principle which areintended to improve the speed and efficiencyof theorem-proving programs based on thisrule of inference. It is proved thxt two ofthe refinements preserve the logicalcompleteness of the proof procedure whenused separately, but not when used incon junction. The results of some preliminaryexperiments with the refinements are given.
Presented at the IRIA Symposium onAutoma,tic Deduction, Versailles, France,December 16-2 1, 1968.
+AIM-82 AD6856 14Zohar Manna, Amir Pneuli,Formalization of Properties of RecursivelyDefined Functions,26 pages, March 1969.
This paper is concerned with therelationship between the convergence,correctness and equivalence of recursivelydefined functions and the satisfiability (orunsatisfiability) of certain first-order formulas.
oAIM-83 cs- 130Roger C. Schank,A Conceptual Representatioll forComputer-oriented Semantics,Thesis: Ph.D. in Ltnguistlcs U. of Texas,201 pages, March 1969.
A. I. MEMO ABSTRACTS 85
the human method for doing so. Someprevious attempts at organizing themachine’s data base conceptually arediscussed. A conceptually-orienteddependency grammar is posited as aninterlingua that may be used as an abstractrepresentation of the underlying conceptualstructure. The conceptual dependencies areutilized as the highest level in a stratifiedsystem that incorporates language-specificrealization rules to map from concepts andtheir relations, into sentences. In order togenerate coherent sentences, a conceptualsemantics is posited that limits possibleconceptu’al dependencies to statements aboutthe system’s knowledge of the real world.This is done by the creation of semantic filesthat serve to spell out the definingcharacteristics of a given concept andenumerate the possibilities for relations withother concepts within the range of conceptualexperience. The semantic files are created, inpart, from a hierarchical organitatlon ofsemantic categories. The semantic category ispart of the definition of a concept and theinformation at the nodes dominating thesemantic category in the hierarchical treemay be used to fill in the semantic file. It ispossible to reverse the realization ‘rules tooperate on sentences and produce aconceptual parse. All potential parses arechecked with the conceptual semantics inorder to eliminate semantic and syntacticambiguities. The system has beenprogrammed; coherent sentences have beengenerated and the parser is operable. Theentire systekhjs posited as a viable linguistictheory.
::~AIM-84David Canfield Smith,MLISP Users’ Manual,57 pages, January 1969.
AD691791
introduction of a more visual flow of controlwith block structure and mnemonic keywords, and language redundancy. Inaddition, some ‘meta-constructs’ areintroduced to increase the power of thelanguage.
oAIM-85 CS- 127 AD687720Pierre Vicens,Aspects of Speech Recognition byComputer,Thesis: Ph.D. in Computer Science,2 10 pages, April 1969.
This thesis describes techniques ’ andmethodology which are useful in achievingclose to real-time recognition of speech bycomputer. To analyze connected . speechutterances, any speech recognition systemmust perform the following processes:preprocessing, segmentation, segmentclassification, recognition of words,recognition of sentences. We presentimplemented solutions to each of theseproblems which achieved accuraterecognition in all the trial cases.
((AIM-86 AD69 1788Patrick J. Hayes,A Machine-oriented Formulation of theExtended Functional Calculus,44 pages, June 1969.
The Extended Functional Calculus (EFC), athree-valued predicate calculus intended as alanguage in which to reason about the resultsof computations, is described in some detail.A formal semantics is given. A machine-oriented (axiomless) inference system forEFC is then described and its completenessrelative to the semantics is proved by themethod of Semantic Trees. Finally someremarks are made on efficiency.
MLISP is a LISP pre-processor designed tofacilitate the writing, use, and understandingof LISP progams. This is accomplishedthrough parentheses reduction, comments,
86 A. I. MEMO ABSTRACTS
:::AIM-87 AD69 1789 ((AIM-90John McCarthy, AI. Project Staff, Anthony C. Hearn,Project Technical Report, Starrdard LISP,98 pages, June 1969. 33 pages, May 1969.
Plans and accomplishments of the StanfordArtificial Intelligence Project are reviewed inseveral areas including: theory (epistemologyand mathematical theory of computation),visual perception and control (Hand-eye andCart), speech recognition by computer,heuristics in machine learning and automaticdeduction, models of cognitive processes(Heuristic DENDRAL), Language Research,and Higher Mental Functions. This is anexcerpt of a proposal to ARPA.
:::AIM-88Roger C. Schank,
AD69 1790
Linguist its from a Conceptual Viewpoint(Aspects of Aspects of a Theory of Syntax),22 pages, April 1969.
Some of the assertions made by Chomsky inAspects of Syntax are considered. Inparticular, the notion of a ‘competence’ modelin linguistics is criticized. Formal postulatesfor a conceputally-based linguistic theory arepresented.
AIM-89 cs- 125 AD692390Jerome A. Feldman, J. Gips, J. J. Horning,and S. Reder,Grammatical Complexity and Inference,100 pages, June 1969.
The problem of inferring a grammar for aset of symbol strings is considered and anumber of new decidability results obtained.Several notions of grammatical complexityand their properties are studied. Thequestion of learning the least complexgrammar for a set of strings is investigatedleading to a variety of positive and negativeresults. This work is part of a continuingeffort to study the problems of representationand generalization through the grammaticalinference question.
AD69 1799
A uniform subset of LISP I.5 capable ofassembly under a wide range of existingcompilers and interpreters is described.
-
AIM-9 1J. A. Campbell and Anthony C. Hearn,Symbolic Analysis of Feynman Diagramsby Computer,73 pages, August 1969.
We describe a system of programs in thelanguage LISP 1.5 which handles all stagesof calculation from the specification of anelementary-particle process in terms of aHamiltonian of interaction or Feynmandiagrams to the derivation of an absolutesquare of the matrix element for the process.Examples of significant parts of the programsare presented in the text, while a detailedlisting of this material is contained in twoAppendices which are available on requestfrom the authors.
((AIM-92Victor D. Scheinman,Design of a Computer ControlledManipulator,Thesis: Eng. in Mechanical Engineering,53 pages, June 1969.
This thesis covers the preliminary systemstudies, the design process, and the designdetails associated with the design of a newcomputer controlled manipulator for theStanford Artificial Intelligence Project. Asystems study of various manipulatorconfigurations, force sensing methods, andsuitable components and hardware was firstperformed. Based on this study, a generaldesign concept was formulated. This conceptWaS then developed into a detailedmanipulator design, having six degrees offreedom, all electric motor powered. The
A. I. MEMO ABSTRACTS 87
manipulator has exceptionally high positionaccuracy, comparatively fast feedback servoperformance, and approximately human armreach and motion properties. Supportingsome of the design details and selections areseveral examples of the design calculationprocedure employed.
AIM-93.1Jerome Feldman,
AD693106
Some Decidability Results 011 GrammaticalInference and Complexity,3 1 pages, August 1969, revised May 1970.
The problem of grammatical -inference isconsidered and a number of positive answersto decidability questions obtained.Conditions are prescribed under which it ispossible for a machine to infer a grammar(or the best grammar) for even the generalrewriting systems.
This revision was motivated by the discoverythat our original definition ofapproachability was too weak and could besatisfied by trivial inference devices.Definition 1.2 and the surrounding materialdiscuss this situation.
. .
The theorems in Section 2 have been slightlyreordered and new proofs given. Theexplicit use of a bounding function gives riseto an important new result, Corollary 2.4.Section 3 is changed primarily in the moredetailed discussion of mixed strategymachines. .
\AIM-94 ’ AD69239 1Kenneth Mark Colby, Lawrence Tesler,Horace Enea,Experiments With a Search Algorithm onthe Data Base of a Human Belief Structure,28 pages, August 1969.
Problems of collecting data regarding humanbeliefs are considered. Representation of thisdata in a computer model designed to judgecredibility involved paraphrasings from
natural language into the symbolicexpressions of the programming languageMLISP. Experiments in processing this datawith a particular search algorithm aredescribed, discussed and criticized.
((AIM-95Zohar Manna,
AD69497 I
The Correctness of Non-deterministicPrograms,44 pages, August 1969.
In this paper we formalize properties of non-deterministic programs by means of thesatisfiability and validity of formulas in first-order logic. Our main purpose is toemphasize the wide variety of possibleapplications of the results.
(<AIM -96 CS- 138Claude Cordell Green,
AD696394
The Application of Theorem Proving toQuestion-answering Systems,Thesis: Ph.D. in Electrical Engineering,166 pages, August 1969.
This paper shows how a question-answeringsystem can use first-order logic as its languageand an automatic theorem prover basedupon the resolution inference principle as itsdeductive mechanism. The resolution proofprocedure is extended to a constructive proof ’procedure. An answer construction algorithmis given whereby the system is able not onlyto produce yes or no answers but also to findor construct an object satisfying a specifiedcondition. A working computer program,QAS, based on these ideas, is described. Theperformance of the program, illustrated byextended examples, compares favorably withseveral other question-answering programs.
Methods are presented for solving statetransformation problems. In addition toquestion-answering, the program can doautomatic programming (simple programwriting, program verifying, and debugging),control and problem solving for a simple
88 A. I. MEMO ABSTRACTS
robot, pattern recognition (scene description),and pu tzles.
AIM-97 AD694972Kenneth Mark Colby, David Canfield Smith,Dialogues Betweell Humans arld anArtificial Belief System,‘28 pages, August 1969.
An artificial belief system capable ofconducting on-line dialogues with humanshas been conntructed. It accepts information,answers questions and establishes acredibility for the information it acquiresand for its human informants. Beginningwith beliefs of high credibility from a highlybelieved source, the system is being subjectedto the experience of dialogues with otherhumans.
AIM-98 cs- 139 AD69540 1James Jay Horning,A Study of Granlmatical Inference,Thesis: Ph.D. in Computer Science,
- 166 pages, August 1969.
The present study has been motivated by thetwin goals of devising useful inferenceprocedures and of demonstrating a soundforma1 basis for such procedures. Theformer has led to the rejection of formallysimple solutions involving restrictions whichare unreasonable in practice; the latter, to therejection of heuristic “bags of tricks” whoseperformance is in general imponderable.Part I states the general grammaticalinference problem for formal languages,reviews previous work, establishes definitionsand notation, and states my position for aparticular class of grammatical inferenceproblems based on an assumed probabilisticstructure. The fundamental results arecontained in Chapter V; the remainingchapters discuss extensions and removal ofrestrictions. Part III covers a variety ofrelated topics, none of which are treated inany depth.
AIM-99Bruce G. Buchanan, C. L. Sutherland, E. A.Feigenbaum,Toward an Understanding of InformationProcesses of Scientific Inference in theCortex t of Organic Chemistry,66 pages, September 1969.
The program called Heuristic DENDRALsolves scientific induction problems of thefollowing type: given the mass spectrum ofan organic molecule, what is the mostplausible hypothesis of organic structure thatwill serve to explain the given empiricaldata. Its problem solving power derives inlarge measure from the vast amount ofchemical knowledge employed in controllingsearch and making evaluations.
A brief description of the task environmentand the program is given in Part I. Recentimprovements in the design of the programand the quality of its performance in thechemical task environment are noted.
The acquisition of task-specific knowledgefrom chemist-‘experts’, the representation ofthis knowledge in a form best suited tofacilitate the problem solving, and the mosteffective deployment of this body ofknowledge in restricting search and makingselections have been major foci of ourresearch. Part II discusses the techniquesused and problems encountered in elicitingmass spectral theory from a cooperativechemist. A sample ‘scenario’ of a session witha chemist is exhibited. Part III discussesmore genera1 issues of the representation ofthe chemical knowledge and the design ofprocesses that utilize it effectively. Theinitial, rather straight-forward,implementations were found to have seriousdefects. These are discussed. Part IV isconcerned with our presently-conceivedsolutions to some of these problems,particularly the rigidity of processes andknowledge-structures.
A. I. MEMO ABSTRACTS 89
The paper concludes with a bibliography ofpublications related to the DENDRAL effort.
AIM-100Zohar Manna, John McCarthy,
Properties of Programs aud PartialFunction Logic,2 I pages, October 1969.
, We consider recursive definitions whichconsist of Algol-like conditional expressions.By specifying a computation rule forevaluating such recursive definition, itdetermines a partial function. However, fordifferent computation rules, the samerecursive definition may determine differentpartial functions. We distinguish betweentwo types of computation rules: sequentialand parallel.
The purpose of this paper is to formalizeproperties (such as termination, correctnessand equivlance) of these partial functions bymeans of the satisfiability or validity ofcer.tain formulas in partial function logic.
This paper was presented in the 5thMachine Intelligence Workshop held atEdinburgh (September 15-20, 1969), and willbe published in Machine Intelligence 3(Edinburgh University Press, 1970).
:::AIM- 101Richard Paul, G. Falk, J. A. Feldman,The Computer Representatiotl of SimplyDescribed Scenes,I6 pages, O.ctober 1969,
This paper describes the computerrepresentation of scenes consisting of anumber of simple three-dimensional objects.One method of representing such scenes is aspace oriented representation whereinformation about a region of space isaccessed by its coordinates. Anotherapproach is to access the information byobject, where, by giving the object name, itsdescription and position are returned.
As the description of an object is lengthy, itis desirable to group similar objects. Groupsof similar objects can be represented in termsof a common part and a number ofindividual parts. If it is necessary tosimulate moving an object then only theindividual information need be saved.
({AIM-102Donald A. Waterman,Generalization Learning for Automatingthe Learning of Heuristics,74 pages, July 1969.
This paper investigates the problem ofimplementing machine learning of heuristics.First, a method of representing heuristics asproduction rules is developed whichfacilitates dynamic manipulation of theheuristics by the program embodying them.Second, procedures are developed whichpermit a problem-solving program employingheuristics in production rule form to learn toimprove its performance by evaluating andmodifying existing heuristics andhypothesizing new ones, either during anexplicit training process or during normalprogram operation. Third, the feasibility ofthese ideas in a complex problem-solvingsituation is demonstrated by using them in aprogram to make the bet decision in drawpoker. Finally, problems which merit furtherinvestigation are discussed, including theproblem of defining the task environmentand the problem of adapting the system toboard games.
AIM-103John Allen, David LuckhamAt1 Interactive Theorem-proving Program,27 pages, October 1969.
We present an outline of the principlefeatures of an on-line mterative theorem-proving program, and a brief account of theresults of some experiments with it. Thisprogram has been used to obtain proofs ofnew mathematical results recently announced
90 A. I. MEMO ABSTRACTS
without proof in the Notices of the AmericanMathematical Society.
AIM-104Joshua Lederberg, Georgia Sutherland, B. G.Buchanan, E. A. Feigenbaum,A Heuristic Program for Solving aScientific Inference Problem: Summary ofMotivation atid Implementation,I5 pages, November 1969.
The primary motivation of the HeuristicDENDR AL project is to study and modelprocesses of inductive inference in science, inparticular, the formation of hypotheses whichbest explain given sets of empirical data.The task chosen for detailed study is organicmolecular structure determination using massspectral data and other associated spectra.This paper first summarizes the motivationand general outline of the approach. Next, asketch is given of how the program worksand how good its performance is at thisstage. The paper concludes with acomprehensive list of publications of theproject.
:::AIM- 105Manfred Heuckel,An Operator Which Locates Edges inDigitized Pictures,37 pages, October 1969.
This paper reports the development of anedge finding operator (subroutine) whichaccepts the digitized light intensities within asmall disc-shaped subarea of a picture andyields a description of any edge (brightnessstep) which may pass over the disc. Inaddition, the operator reports a measure ofthe edge’s reliability. A theoretical effortdisclosed the unique best operator whichsatisfies a certain set of criteria for a localedge recognizer. The main concerns of thecriteria are speed and reliability in thepresence of noise.
AIM- 106Michael Edwin Kahn,The Near-minimum-time Control of Open-loop Articulated Riueinatic Chains,Thesis: Ph.D. in Mechanical Engineering,I 7 I pages, December 1969.
The time-optimal control of a system of rigidbodies connected in series by single-degree-of-freedom joints is studied.
The dynamical equations of the system arehighly nonlinear and a closed-formrepresentation of the minimum-time feedbackcontrol is not possible. However, asuboptimal feedback control which providesa close approximation to the optimal controlis developed.
The suboptimal control is expressed in termsof switching curves for each of the systemcontrols. These curves are obtained from thelinearized equations of motion for the system.Approximations are made for the effects ofgravity loads and angular velocity terms inthe nonlinear equations of motion.
Digital simulation is used to obtain acomparison of response times of the optimaland suboptimal controls. The speed ofresponse of the suboptimal control is foundto compare quite favorably with the responsespeed of the optimal control.
The analysis is applied to the control ofthree joints of a mechanical manipulator.Modifications of the suboptimal control foruse in a sampled-data system are shown toresult in good performance of a hydraulicmanipulator under computer control.
?:cAIM- 107Gilbert Falk,Soine Implicatioris of Planarity forMachiue Perception,27 pages, December 1969.
The problem of determining the shape and
A. I. MEMO ABSTRACTS
orientation of an object based on one ormore two-dimensional images is considered.For a restricted class of projections it isshown that monocular information is often“nearly” sufficient for complete specificationof the object viewed.
AIM-108Michael D. Kelly,Edge Detection in Pictures by ComputerUsing Planning,28 pages, January 1970.
This paper describes a program forextracting an accurate outline of a man’shead from a digital picture. The programaccepts as input digital, grey scale picturescontaining people standing in front ofvarious backgrounds. The output of theprogram is an ordered list of the pointswhich form the outline of the head. Theedges of background objects and the interiordetails of the head have been suppressed.
The program is successful because of animproved method for edge detection whichuses heuristic planning, a technique drawnfrom artificial intelligence research inproblem solving. In brief, edge detectionusing planning consists of three steps. A newdigital picture is prepared from the original;the ne,w picture is smaller and has less detail.Edges of objects are located in the reducedpicture. The edges found in the reducedpicture are used as a plan for finding edgesin the original picture.
‘\\::tAIM- 109Roger C. Schank, Lawrence Tesler, Sylvia *Weber,Spinoza If: Conceptual Case-based Natural-Language Analysis,107 pages, January 1970.
This paper presents the theoretical changesthat have developed in ConceptualDependenty Theory and their ramificationsin computer analysis of natural language.
-
91
The major items of concern are: theelimination of reliance on ‘grammar rules’ forparsing with the emphasis given t oconceptual rule based parsing, thedevelopment of a conceptual case system toaccount for the power of conceptualizations;the categorization of ACT’s based onpermissible conceptual cases and othercriteria. These items are developed anddiscussed in the context of a more powerfulconceptual parser and a theory of languageunderstanding.
ti{AIM-110Edward Ashcroft, Zohar Manna,Formalization of Properties of ParallelPrograms,58 pages, February 1970.
In this paper we describe a class of parallelprograms and give a formalization of certainproperties of such programs in predicatecalculus.
Although our programs are syntacticallysimple, they do exhibit interaction betweenasynchronous parallel processes, which is theessential feature we wish to consider. Theformalization can easily be extended to morecomplicated programs.
Also presented is a method of simplifyingparallel programs, i.e., constructing simplerequivalent programs, b a s e d o n the“independence” of statements in them. Withthese simplications our formalization gives apractical method for proving properties ofsuch programs.
GAIM- 111Zohar Manna,Second-order Mathematical Theory ofComputation,25 pages, March 1970.
In this work we show that it is possible toformalize all properties regularly observed in(deterministic and non-deterministic)algorithms in second-order predicate calculus.
92
Moreover, we show that for any givenalgorithm it suffices to know how toformalize its “partial correctness” by asecond-order formula in order to formalizeall other properties by second-order formulas.
This result is of special interest since “partialcorrectness” has already been formalized insecond-order predicate calculus for manyclasses of algorithms.
This paper will be presented at the ACMSymposium on Theory of Computing (May1970). 5
AIM-1 12Franklin D. Hilf, Kenneth M. Colby, DavidC. Smith, William K. Wittner,Machine-mediated Interviewing,27 pages, March 19’70.
A technique of psychiatric interviewing isdescribed in which patient and interviewercommunicate by means of remotely locatedIeletypes. Advantages of non-nonverbalcommunication in the study of thepsychiatric interview and in the developmentof a computer program designed to conductpsychiatric interviews are discussed.Transcripts from representative interviewsare reproduced.
:>AIM-113.Kenneth Mark Colby, Franklin D. Hilf,William A. Hall,A Mute Patient’s Experience WithMachine-mediated Interviewing, -19 pages, March 1970.
A hospitalized mute patient participated inseven machine-mediated interviews, excerptsof which are presented. After the fifthinterview he began to use spoken languagefor communication. This novel technique issuggested for patients who are unable toparticipate in the usual vis-a-vis interview.
t4AIM- 114Alan W. Biermann, Jerome A. Feldman,On the Synthesis of Finite-state Acceptors,3 1 pages, April 1970.
Two algorithms are presented for solving thefollowing problem: Given a finite-set S ofstrings of symbols, find a finite-state machinewhich will accept the strings of S andpossibly some additional strings which“resemble” those of S. The approach used isto directly construct the states and transitionsof the acceptor machine from the stringinformation. The algorithms include aparameter which enable one to increase theexactness of the resulting machine’s behavioras much as desired by increasing the numberof states in the machine. The properties ofthe algorithms are presented and illustratedwith a number of examples.
The paper gives a method for identifying afinite-state language from a randomly chosenfinite subset of the language if the subset islarge enough and if a bound is known onthe number of states required to recognizethe language. Finally, we discuss some of theuses of the algorithms and their relationshipto the problem of grammatical inference.
AIM-l 15Ugo Montanari,On the Optimal Detection of Curves inNoisy Pictures,35 pages, March 1970.
A technique for recognizing systems of linesis presented, in which the heuristic of theproblem is not embedded in the recognitionalgorithm but is expressed in a figure ofmerit. A multistage decision process is thenable to recognize in the input picture theoptimal system of lines according to the givenfigure of merit. Due to the global approach,greater flexibility and adequacy in theparticular problem is achieved. The relationbetween the structure of the figure of meritand the complexity of the optimization
A. I. MEMO ABSTRACTS
process is then discussed. The methoddescribed is suitable for parallel processingbecause the operations relative to each statecan be computed in parallel, and the numberof stages is equal to the length N of thecurves (or to log2(N) if an approximatemethod is used).
:sAIM- I I6Kenneth Mark Colby,Mind and Brain, Again,10 pages, March 1970.
Classical , mind-brain questions appeardeviant through the lens of an analogycomparing mental processes withcomputational processes. Problems ofreducibility and personal consciousness arealso considered in the light of this analogy.
:::AIM- 117John McCarthy, et al,Project Technical Report,75 pages, April 1970.
Current research is reviewed in artificialintelligence and related areas, includingrepresentation theory, mathematical theory ofcomputation, models of cognitive processes,speech recognition, and computer vision.
AIM- 118Ugo Montanari,Heuristically Guided Search andChromosotne Matchiag,29 pages, April 1970.
Heuristically guided search is a techniquewhich takes systematically into accountinformation from the problem domain fordirecting the search. The problem is to findthe shortest path in a weighted graph from astart vertex Va to a goal vertex Vz: for everyintermediate vertex, an estimate is availableof the distance to Vz. If this estimatesatisfies a consistency assumption, analgorithm by Hart, Nilsson and Raphael isguaranteed to find the optimum, looking at
-
93
the a priori minimum number of vertices. Inthis paper, a version of the above algorithmis presented, which is guaranteed to succeedwith the minimum amount of storage. Anapplication of this technique to thechromosome matching problem is thenshown. Matching is the last stage ofautomatic chromosome. analysis procedures,and can also solve ambiguities in theclassification stage. Some peculiarities of thiskind of data suggest the use of a nheuristically guided search algorithm insteadof the standard Edmonds’ algorithm. Themethod that we obtain in this way is provedto exploit the clustering of chromosome data:a h-near-quadratic dependence from thenumber of chromosomes is obtained forperfectly clustered data. Finally, someexperimental results are given.
::cAIM- 119Joseph Becker,An Illformation-processing Model ofIntermediate-Level Cognition,123 pages, May 1970.
There is a large class of cognitive operationsin which an organism adapts its previousexperience in order to respond properly to anew situation -- for example: the perceptualrecognition of objects and events, theprediction of the immediate future (e.g. intracking a moving object), and theemployment of sensory-motor “skills”. Takenall together, these highly efficient processesform a cognitive subsystem which is-intermediate between the low-level sensory-motor operations and the more deliberateprocesses of high-level ‘thought’.
The present report describes a formalinformation-processing model of this‘Intermediate-Level’ cognitive system. Themodel includes memory structures for thestorage of experience, and processes forresponding to new events on the basis ofprevious experience. In addition, theproposed system contains a large number of
94
mechanisms for making t h e response-selection process highly efficient, in spite ofthe vast amount of stored information thatthe system must cope with. These devicesinclude procedures for heuristicallyevaluating alternative su bprocesses, forguiding the search through memory, and forreorganizing the information in memory intomore efficient representations.
:::AIM- 120Kenneth Mark Colby, David Canfield Smith, ~AIlvf- 122Computer as Catalyst in the Treatment of Roger C. Schank,Nonspeaking Autistic Children, ‘Semantics’ in Conceptual Analysis,32 pages, April 1970. 5fi pages, May 1970.
Continued experience with a computer-aidedtreatment method for nonspeaking autisticchildren has demonstrated improvementeffects in thirteen out of a series of seventeencases. Justification for this conclusion isdiscussed in detail. Adoption of this methodby other research groups is needed for thefuture development of computer-aidedtreatment.
This paper examines the question of what asemantic theory should account for. Someaspects of the work of Katz, Fillmore, Lakoffand Chomsky are discussed. Semantics isconcluded to be the representation problemwith respect to conceptual analysis. Thebeginnings of a solution to this problem arepresented in the light of developments inconceptual dependency theory.
::<AIM-121Irwin Sobel,Camera Models arld Machine Perception,Thesis: Ph.D. in Electrical Engineering,89 pages, May 1970.
We have developed a parametric model for acomputer-controlled moveable camera on apan-tilt head. The model expresses thetransform relating object space to imagespace as a function of the control variables -of the camera. We constructed a calibrationsystem for measuring the model parameterswhich has a demonstrated accuracy morethan adequate for our present needs. Wehave also identified the major source of errorin mode1 measurement to be undesired imagemotion and have developed means ofmeasuring and compensating for some of itand eliminating other parts of it. The systemcan measure systematic image distortions ifthey become the major accuracy limitation.
We have shown how to generalize the modelto handle small systematic errors due toaspects of pan-tilt head geometry notpresently accounted for.
We have demonstrated the model’sapplication in stereo vision and have shownhow it can be applied as a predictive devicein locating objects of interest and centeringthem in an image.
AIM-123Bruce G. Buchanan, Thomas E. Headrick,Some Speculation About ArtificialIntelligence and Legal Reasoning,54 pages, May 1970.
Legal reasoning is viewed here as a complexproblem-solving task to which the techniquesof artificial intelligence programming may beapplied. Some existing programs arediscussed which successfully attack variousaspects of the problem, in this and other taskdomains. It remains an open question, to beanswered by intensive research, whethercomputers can be programmed to do creativelegal reasoning. Regardless of the answer, itis argued that much will be gained by theresearch.
A. I. MEMO ABSTRACTS
AIM-124M. M. Astrahan,Speech Analysis by Clustering, or theHyperphollelne Method,22 pages, June 1970.
In this work, measured speech waveformdata was used as a basis for partitioning anutterance into segments and for classifyingthose segments. Mathematical classificationswere used instead of the traditionalphonemes or linguistic categories. Thisinvolved clustering methods applied toh yperspace points representing periodicsamples of speech waveforms. The clustercenters, or hyperphonemes (HPs), were usedto classify the sample points by the nearest-neighbor technique. Speech segments wereformed by grouping adjacent points with thesame classification. A dictionary of 54different words from a single speaker wasprocessed by this method. 216 utterances,representing four more repetitions by thesame speaker each of the original 54 words,were similarly analyzed into strings ofhyperphonemes and matched against thedictionary by heuristically developedformulas. 87% were correctly recognized,although almost no attempt was made tomodify and improve the initial methods andparameters.
:::AIM- I25Kenneth Mark Colby, Sylvia Weber,Franklin Hilf,Artificial Paranoia,35 pages, July 1970.
A case of artificial paranoia has beensynthesized in the form of a computer model.Using the test operations of a teletypedpsychiatric interview, clinicians judge theinput-output behavior of the model to beparanoid. Forma1 validation of the modelwill require experiments involvingindistinguishability tests.
95
AIM- 126 CS- I69 AD71 1329Donald E. Knuth,Exatnples of Formal Semantics,34 pages, July 1970.
A technique of formal definition, based onrelations between ‘attributes’ associated withnonterminal symbols in a context-freegrammar, is illustrated by severalapplications to simple yet typical problems.First we define the basic properties of lambdaexpressions, involving substitution andrenaming of bound variables. Then a simpleprogramming language is defined usingseveral different points of view. Theemphasis is on ‘declarative’ rather than‘imperative’ forms of definition.
AIM- 127 cs- 174 AD71 1395Zohar Manna, Richard J. Waldinger,Towards Automatic Program Synthesis,54 pages, July 1970.
An elementary outline of the theorem-proving approach to automatic programsynthesis is given, without dwelling ontechnical details. The method is illustratedby the automatic construction of bothrecursive and iterative programs operatingon natural numbers, lists, and trees.
In order to construct a program satisfyingcertain specifications, a theorem induced bythose specifications is proved, and the desiredprogram is extracted from the proof. Thesame technique is applied to transformrecursively defined functions into iterativeprograms, frequently with a major gain inefficiency.
It is emphasized that in order to construct aprogram with loops or with recursion, theprinciple of mathematical induction must beapplied. The relation between the version ofthe induction rule used and the form of theprogram constructed is explored in somedetail.
96 A. I. MEMO ABSTRACTS
AIM-128 cs- 166 AD713841Erik J. Sandewall,Representing Natural-languageInfomatiolr in Predicate Calculus,27 pages, July 1970.
A set of general conventions are proposed forrepresenting natural language information inman y-sorted first order predicate calculus.The purpose is to provide a testing-groundfor existing theorem-proving programs.
:::AlM- 129 CS- 167 AD712460S higeru Igarashi,Siemantics of ALGOL-like Statements,95 pages, June 1970.
The semantics of elementary Algol-likestatements is discussed, mainly based on anaxiomatic method.
Firstly, a class of Algol-like statements isintroduced by generalized inductivedefinition, and the interpretation of theslatements belonging to it is defined in theform of a function over this class, using theinduction principle induced by the abovedefinition. Then a category of program isintroduced in order to clarify the concept ofequivalence of statements, which becomes aspecial case of isomorphism in that category.
.A revised formal system representing theconcept of equivalence of Algol-ii kestatements is presented, followed byplemen tary metatheorems.
Finally, a process of decomposition of Algol-like statements, which can be regarded as aconceptual compiler, or a constructivedescription of semantics based on primitiveactlons, is defined and its correctness isproved formally, by the help of the inducedinduction principle.
AIM-130 cs- 168 AD7 13252Michael D. Kelly,Visual Identification of People byComputer,Thesis: Ph.D. in Computer Science,238 pages, July 1970.
This thesis describes a computer programwhich performs a complex picture processingtask. The task is to choose, from a collectionof pictures of people taken by a TV camera,those pictures that depict the same person.The primary purpose of this research hasbeen directed toward the development of newtechniques for picture processing.
In brief, the program works by finding thelocation of features such as eyes, nose, orshoulders in the pictures. Individuals areclassified by measurements between suchfeatures. The interesting and difficult part ofthe work reported in this thesis is t h edetection of those features in digital pictures.The nearest neighbor method is used foridentification of individuals once a set ofmeasurements has been obtained.
The success of the program is due to andillustrates the heuristic use of context andstructure. A new, widely useful, techniquecalled planning has been applied to pictureprocessing. Planning is a term which isdrawn from artificial intelligence research inproblem solving.
The principal positive result of this researchis the use of goal-directed techniques tosuccessfully locate features in cluttered digitalpictures. This success has been verified bydisplaying the results of the feature findingalgorithms and comparing these locationswith the locations obtained by hand fromdigital printouts of the pictures. Successfulperformance in the task of identification ofpeople provides further verification for thefeature finding algorithms.
,
A. I. MEMO ABSTRACTS 97
AIM-131 CS- 176 AD715128Edward A. Feigenbaum, Bruce C. Buchanan,Joshua Lederberg,011 Generality alrd Problem Solving: a CaseStudy Using the Dendral Program,48 pages, August 1970.
S-space in addition to complete monocularinformation are sufficient to specify all thevisible point locations precisely.
QAIM- 133 CS-181Anthony C. Hearn,Reduce 2,
Heuristic DENDRAL is a computer programwritten to solve problems of inductiveinference in organic chemistry. This paperwill use the design of Heuristic DENDRALand its performance on different problemsfor a discussion of the following topics:
1. the design for generality;2. the performance problems attendent upon
too much generality
Diskfile: REDUCE.ACH[AIM,DOCI,85 pages, October 1970.
3. the coupling of expertise to the generalproblem solving processes,
4. the symbiotic relationship betweengenerality and expertnness of problemsolving systems.
This manual provides the user with adescription of the algebraic programmingsystem REDUCE 2. The capabilities of thissystem include:
I) Expansion and ordering of rational-functions of polynomials,
2) symbolic differentiation of rationalfunctions of polynomials and generalfunctions,
We conclude the paper with a view of thedesign for a general problem solver that is avariant of the “big switch” theory ofgenerality.
:::AIM- 132 CS- 180 AD715665Gilbert Falk,Cowputer Iiiterpretatioii of Imperfect LineData as a Three-dimensiorral Scene,Thesis: Ph.D. in Electrical Engineering,187 pages, August 1970.
3) substitutions and pattern matching in awide variety of forms,
4) calculation of the greatest commondivisor of two polynomials,
5) automatic and user controlledsimplification of expressions,
6) calculations with symbolic matrices,7) a complete language for symbolic
calculations, in which the REDUCEprogram itself is written,
8) calculations of interest to high energyphysicists including spin l/2 and spin 1algebra,
9) tensor operations.
The major portion of this paper describes aheuristic scene description program. Thisprogram accepts as input a scene representedas a line drawing. Based on a set of knownobject models the program attempts todetermine the identity and location of eachobject viewed. The most significant featureof the program is its ability to deal withimperfect input data.
({AIM- 134 CS- 182Jay Martin Tenenbaum,
AD748565
Accommodation III Computer Vision,Thesis: Ph.D. in Electrical Engineering,452 pages, September 1970.
We also present some preliminary resultsconcerning constraints in projections ofplanar-faced solids. We show that for arestricted class of projects, 4 points located in
We describe an evolving computer visionsystem in which the parameters of thecamera are controlled by the computer. It isdistinguished from conventional pictureprocessing systems by the fact that sensoraccommodation is automatic and treated asan integral part of the recognition process.
-
98 A. I. MEMO ABSTRACTS
A machine, like a person, comes in contactwith far more visual information than it canprocess. Furthermore, no physical sensor cansimultaneously provide information aboutthe full range of the environment.Consequently, both man and machine mustaccommodate their sensors to emphasizeselected characteristics of the environment.
Accommodation improves the reliability andefficiency of machine perception by matchingthe information provided by the sensor withthat required by specific perceptual functions.T h e <advantages of accommodation aredemonstrated in the context of five keyfunctions in computer vision: acquisition,contour following, verifying the presence ofan expected edge, range-finding, and colorrecognition.
We have modeled the interaction of camerapara meters with scene characteristics todetermine the composition of an image.Using a priori knowledge of theenvironment, the camera is tuned to satisfythe information requirements of a particulartask.
Task performance depends implicitly on theappropriateness of available information. Ifa function fails to perform as expected, andif this failure is attributable to a specificimage deficiency, then the relevantaccommodation parameters can be refined.
This schema for automating sensoraccommodation can be applied in a-varietyof perceptual domains.
MLISP is a high level list-processing andsymbol-manipulation language based on theprogramming language LISP. MLISP
programs are translated into LISP programsand then executed or compiled. MLISPexists for two purposes: (I) to facilitate thewriting and understanding o f L I S Pprograms; (2) to remedy certain importantdeficiencies in the list-processing ability ofLISP.
-
<<AIM- 136 CS- 183 AD ‘7 17600George M. White,Machine Learning Through SignatureTrees. Applications to Human Speech,40 pages, October 1970.
Signature tree ‘machine learning’, patternrecognition heuristics are investigated for thespecific problem of computer recognition ofhuman speech. When the data base of givenutterances is insufficient to establish trendswith confidence, a large number of featureextractors must be employed and ‘recognition’of an unknown pattern made by comparingits feature values with those of knownpatterns. When the data base is replete, a‘signature’ tree can be constructed andrecognition can be achieved by t h eevaluation of a select few features. Learningresults from selecting an optimal minimal setof features to achieve recognition. Propertiesof signature trees and the heuristics for thistype of learning are of primary interest inthis exposition.
::{AIM- 137Donald E. Knuth,An Empirical Study of Fortran in Use,44 pages, November 1970.
A sample of programs, written in Fortran bya wide variety of people for a wide variety ofapplications, was chosen ‘at random’ in anattempt to discover quantitatively ‘whatprogrammers really do’. Statistical results ofthis survey are presented here, together withsome of their apparent implications forfuture work in compiler design. Theprinciple conclusion which may be drawn isthe importance of a program ‘profile’, namely
A. I. MEMO ABSTRACTS
a table of frequency counts which recordhow often each statement is performed in atypical run: there are strong indications thatprofile-keeping should become a standardpractice in all computer systems, for casualusers as well as system programmers. Somenew approaches to compiler optimization arealso suggested. This paper is the report of athree month study undertaken by the authorand about a dozen students andrepresentatives of the software industryduring the summer of 1970.
AIM-138 cs- 188 PB197161Edward Ashcroft, Zohar Manna,The Translatiou of ‘CO-TO’ Programs to‘WHILE’ Programs,28 pages, January 1971.
In this paper we show that every flowchartprogram can be written without go-tostatements by using while statements. Themain idea is to introduce new variables topreserve the values of certain variables atparticular points in the program; oralternatively, to introduce special, booleanvariables to keep information about thecourse of the computation.
The while programs produced yield the samefinal results as the original flowchart programbut need not perform computations in exactlythe same way. However, the new programspreserve the topology of the originalflowchart program, and are of the same orderof efficiency. .
We also show that this cannot be done ingeneral without adding variables.
AIM-139 CS- 189 AD717601Zohar Manna,Mathematical Theory of PartialCorrect uess,24 pages, January 1971.
In this work we show that it is possible toexpress most properties regu larly observed in
99
algorithms in terms of partial correctness (i.e.,the property that the final results of thealgorithm, if any, satisfy some given input-output relation).
This result is of special interest since partialcorrectness has already been formulated inpredicate calculus and in partial functionlogic for many classes of algorithms.
::cAIM- 140 cs-193Roger C. Schank,Iutentiolr, Memory, aud ComputerUnderstanding,59 pages, January 197 1.
Procedures are described for discovering theintention of a speaker by relating theConceptual Dependence representation of thespeaker’s utterance to the computer’s worldmodel such that simple implications can bemade. These procedures function at levelshigher than that of structure of the memory.Computer understanding of natural languageis shown to consist of the following parts:assigning a conceptual representation to aninput; relating that representation to thememory such as to extract the intention ofthe speaker; and selecting the correctresponse type triggered by such an utteranceaccording to the situation.
<(AIM-141 CS-203 AD730506Bruce G. Buchanan, Joshua Lederberg,The Heuristic DENDRAL Program forExplairlillg Empirical Data,20 pages, February 1971.
The Heuristic DENDRAL program uses aninformation processing model of scientificreasoning to explain experimental data inorganic chemistry. This report summarizesthe organization and results of the programfor computer scientists. The program isdivided into three main parts: planning,structure generation, and evaluation.
The planning phase infers constraints on the
100 A. I. MEMO A&TRACTS
search space from the empirical data input tothe system. The structure generation phasesearches a tree whose termini are models ofchemical models using pruning heuristics ofvarious kinds. The evaluation phase teststhe candidate structures against the originaldata. Results of the program’s analyses ofsome tests are discussed.
AIM-142 CS-205 AD731383Robin Milner,An Algebraic Definitioll of ShulatiorlBetweeu Programs,?I pages, February 1971.
,4 simulation relation between programs ISdefined which is quast-ordering. Mutualsimulation is then an equivalence relation,arid by dividing out by it we abstract from aI‘rogram such details as how the sequencingis controlled and how data is represented.The equivalence classes are approximationsto the algorithms which are realized, orexpressed, by their member programs.
4 technique is given and illustrated forproving simulation and equivalence ofprograms; there is an analogy with Floyd’stechnique for proving correctness ofprograms. Finally, necessary and sufficientconditions for simulation are given.
~a41M 143 cs-209John McCarthy, et al,Project Technical Report,80 pages, March 1971.
AD724867
411 overvlew is presented of current research:” Stanford in artificial intelligence andheuristic programming. This report islargely the text of a proposal to theAdvanced Research Projects Agency for fiscalyears 19’72-3.
AIM- 144 cs-2 19Lynn H. Quam,Computer Coinparisori of Pictures,Thesis: Ph.D. in Computer Science,120 pages, May 1971.
This dissertation reports the development ofdigital computer techniques for detectingchanges in scenes by normalizing andcomparing pictures which were taken fromdifferent camera positions and underdifferent conditions of illumination. Thepictures are first geometrlcally normahzed toa common point of view. Then they arephptometrically normalized to eliminate thedifferences due to diff‘erpnt illumination,camera characteristics, and reflectanceproperties of the scene due to different sunand view angles. These pictures are thengeometrically registered by maximizing thecross correlation between areas in them. Thefinal normalized and registered pictures arethen differenced point by point.
The geometric normalization techniquesrequire relatively accurate geometric modelsfor the camera and the scene, and staticspatial features must be present in thepictures to allow precise geometric alignmentusing the technique of cross correlationmaximization.
PhotometrIc normalization also requires arelatively accurate model for the photometricresponse of the camera, a reflectance modeifor the scene (reflectance as a function of theillumination view, and phase angles) andsome assumptions about the kmds ofreflectance changes which are to be detected.
These techniques have been incorporated ma system for comparmg Mariner 197 1pictures of Mars to detect variable surfacephenomena as well as color and polarizationdifferences. The system has been tested usingMariner 6 and 7 pictures of Mars.
Although the techniques described in this
A. I. MEMO ABSTRACTS 101
dissertation were developed for Marspictures, their use is not limited to thisapplication. Various parts of this softwarepackage, which was developed for interactiveuse on the time-sharing system of theStanford Artificial Tntelligence Project, arecurrently being applied to other scenery.
GAIM- 147 CS-2 16Robert E. Kling,
AD732457
Reasouiug by Analogy with Applicatiomto Heuristic Problem Solving: a Case Study,Thesis: Ph.D. in Computer Science,191 pages, August 1971.
:::AIM- 145 cs-22 1 AD731729Bruce G. Buchanan, Edward A. Feigenbaum,Joshua Lederberg,A Heuristic Programming Study of TheoryFormation in Science,41 pages,, June 1971.
T h e Meta-DENDR AL p rogram i s a avehicle for studying problems of theoryformation in science. The general strategy ofMeta-DENDRAL is to reason from data toplausible generalizations and then toorganize the generalizations into a unifiedtheory. Three main subproblems arediscussed: (1) explain the experimental datafor each individual chemical structure, (2)generalize the results from each structure toall structures, and (3) organize thegeneralizations into a unified theory. Theprogram is built upon the concepts andprogrammed routines already available inthe Heuristic DENDRAL performanceprogram, but goes beyond the performanceprogram in attempting to formulate thetheory which the performance program willuse.
An information-processing approach toreasoning by analogy is developed thatpromises to increase the efficiency of heuristicdeductive problem-solving systems. When adeductive problem-solving system accesses alarge set of axioms more than sufficient for aparticular problem, it will often create manyirrelevant deductions that saturate thememory of the problem solver.
Here, an analogy with some previouslysolved problem and a new unsolved problemis used to restrict the data base to a small setof appropriate axioms. This procedure(ZORBA) is studied in detail for a resolutiontheorem proving system. A set of algorithms(ZORBA-I) which automatically generates ananalogy between a new unproved theorem,TS-A, and a previously proved theorem, T, isdescribed in detail. ZORBA-I is implementedin LISP on a PDP- 10.
AIM-146 CS-224Andrei P. Ershov,Parallel Program miag,-14 pages, July 1971.
PB212183
ZORBA-I is examined in terms of itsempirical performance on parts of analogoustheorems drawn from abstract algebra.Analytical studies are included which showthat ZORBA-I can be useful to aid automatictheorem proving in many pragmatic caseswhile it may be a detriment in certainspecially contrived cases.
This report is based on lectures given atStanford University by Dr. Ershov inNovember, 1970.
AIM- 148 CS-217 AD731730Edward Ashcroft, Zohar Manna, AmirPneuli,Decidable Properties of MonadicFutlctional Schemas,10 pages, July 1971.
We define a class of (monadic) functionalschemas which properly include ‘Ianov’flowchart schemas. We show that the
102 A. I. MEMO ABSTRACTS
termination, divergence and freedomproblems for functional schemas aredecidable. Although it is possible to translatea large class of non-free functional schemasinto equivalent free functional schemas, weshow that this cannot be done in general.We show also that the equivalence problemfor free functional schemas is decidable.Most of the results are obtained from well-known results in Formal Languages andAutomata Theory.
AIM-149 cs-23 I AD 732644Rodney .Albert Schmidt, Jr.,A Study of the Real-time Coritrol of aComputer-driven Vehicle,Thesis: Ph.D. in Electrical Engineering,180 pages, August 1971.
Vehicle control by the computer analysis ofvisual images is investigated. The areas ofguidance, navigation, and incident avoidanceare considered. A television camera is usedas the prime source of visual image data.
1 II the guidance system developed for anexperimental vehicle, visual data is used togain information about the vehicle systemdynamics, as well as to guide the vehicle.This information is used in real time toImprove performance of the non-linear, time-varying vehicle system.
4 scheme for navigation by pilotage throughthe recognition of two dimensional scenes isdeveloped. A method is proposed to link thisto a computer-modelled map in order to _make Journeys.
Various difficulties in avoiding anomolous1 nciden ts in the automatic control ofautomobiles are discussed, together withsuggestions for the application of this studyto remote exploration vehicles or industrialautomation.
*AIM- 150Robert W. Floyd,Toward Interactive Design of CorrectProgram 5,12 pages, September 1971.
We propose an interactive system provingthe correctness of a program, or locatingerrors, as the program is designed.
:::AIM- 15 1 cs-240 AD738568Ralph L. London,Correctness of Two Compilers for a LISPSubset,4 1 pages, October 197 1.
Using mainly structural induction, proofs ofcorrectness of each of two running Lispcompilers for the PDP- 10 computer aregiven. Included are the rationale forpresenting these proofs, a discussion of theproofs, and the changes needed to the secondcompiler to complete its proof.
:::AIM- 152 CS-24 1 AD 732642Alan W. Biermann,On the Iufereuce of Turing Machines fromSample Computations,3 t pages, October 197 1.
An algorithm is presented which when givena complete description of a set of Turingmachine computations f inds a Tur ingma.chine which is capable of doing thosecomputations. This algorithm can serve asthe basis for designing a trainable devicewhich can be trained to simulate any Turingmachine by being led through a series ofsample computations done by that machine,A number of examples illustrate the use ofthe technique and the possibility of theapplication to other types of problems.
::tAIM- 153 CS-242 AD738569Patrick J. Hayes,The Frame Problem and Related Problemsin Artificial Intelligence,18 pages, November 1971.
A. I. MEMO ABSTRACTS 103
r
The frame problem arises in considering thelogical structure of a robot’s beliefs. It hasbeen known for some years, but only recentlyhas much progress been made. The problem
i s described and discussed. Varioussuggested methods for its solution areoutlined, and described fn a un i fo rmnotation. Finally, brief consideration is givento the problem of adjusting a belief system inthe face of evidence which contradicts beliefs.It is shown that a variation on the situationnotation of (McCarthy and Hayes, 1969)permits an elegant approach, and relates thisproblem to the frame problem.
::cAIM- 154 CS-24 3 AD738570Zohar Manna, Stephen Ness, Jean Vuillemin,Inductive Methods for Provilig Propertiesof Programs,24 pages, November 1971.
We have two main purposes in this paper.First, we clarify and extend known resultsabout computation of recursive programs,ernp h asizing the difference between thetheoretical and practical approaches.Secondly, we present and examine variousknown methods for proving properties ofrecursive programs. We discuss in detail twopowerful inductive methods, computationalinduction and structural induction,illustrating their applications by variousexamples. We also briefly discuss some otherrelated methods.
Our aim in this work is to introduceinductive methods to as wide a class ofreaders as possible and to demonstrate theirpower as practical techniques. We ask theforgiveness of our more theoretical-mindedcolleagues for our occasional choice of clarityover precision.
AIM-155- CS-245Jonathan Leonard Ryder,Heuristic Analysis of Large Trees asGenerated in the came of co,Thesis: Ph.D. in Computer Science,300 pages, December 1971.
The Japanese game of Go is of interest bothas a problem in mathematical repesentationand as a game which generates a move treewith an extraordinarily high branchingfactor (100 to 300 branches per ply). Thecomplexity of Go (and the difficulty of Gofor human players) is thought to beconsiderably greater than that of chess. Theconstraints of being able to play a completegame and of being able to produce a movewith a moderate amount of processing timewere placed on the solution.
The basic approach used was to findmethods for isolating and exploring severalsorts of relevant subsections of the globalgame tree. This process depended heavily onthe ability to define and manipulate theentities of Go as recursive functions ratherthan as patterns of stones. A generalmachine-accessible theory of Go wasdeveloped to provide context for programevaluations.
A program for playing Go is now availableon the Stanford PDP- 10 computer. It willplay a complete game, taking about 10 to 30seconds for an average move. The quality ofplay is better than that of a beginner inmany respects, but incompletenesses in thecurrent machine-representable theory of Goprevent the present program from becominga strong player.
AIM- 156 CS-246 AD740141Kenneth Mark Colby, Franklin D. Hilf,Sylvia Weber, Helena C. Kraemer,
- A Resemblance Test for the Validation of aComputer Simulation of ParanoidProcesses,29 pages, November 1971.
A computer simulation of paranoid processesin the form of a dialogue algorithm wassubjected to a validation study using anexperimental resemblance test in whichjudges rate degrees of paranoia present ininitial psychiatric interviews of both
.
104 A. I. MEMO ABSTRACTS
paranoid patients and of versions of theparanoid model. The statistical resultsindicate a satisfactory degree of resemblancebetween the two groups of interviews. It isconcluded that the model provides asuccessful simulation of naturally occuringparanoid processes.
:::AIM-157 cs-247Yorick Wilks,OIle Small Head -- Some Remarks 011 theuse of ‘Model’ in Linguistics,17 pages, December 1971.
I argue that the present situation in formallinguistics, where much new work ispresented as being a “model of the brain”, orof “human language behavior”, is anundesirable one. My reason for thisjudgement is not the conservative(Braithwaitian) one that the entities inquestion are not really models but theories.It is rather that they are called modelsbecause they cannot be theories of the brainat the present stage of brain research, andhence that the use of “model” in this contextis not so much aspirational as resigned aboutour total ignorance of how the brain storesand processes linguistic information. Thereason such explanatory entities cannot betheories is that this ignorance precludes any“semantic ascent” up the theory; i.e.,interpreting the items of the theory in termsof observables. And the brain items,whatever they may be, are not, as Chomskyhas sometimes claimed, in the same positionas the “occult entities” of Physics likeGravitation; for the brain items are nottheoretically unreachable, merely unreached.
I then examine two possible alternate viewsof what linguistic theories should beproffered as theories of: theories of sets ofsentences, and theories of a particular class ofalgorithms. I argue for a form of the latterview, and that its acceptance would also havethe effect of making ComputationalLinguistics a central part of Linguistics,rather than the poor relation it is now.
I examine a distinction among ‘linguisticmodels’ proposed recently by Mey. who wasalso arguing for the self-sufficiency ofComputational Linguistics, though as a‘theory of performance’. I argue that hisdistinction is a bad one, partly for thereasons developed above and partly becausehe attempts to tie it to Chomsky’s inscrutablecompetance-performance dis t inc t ion . Iconclude that the independence and self-sufficiency of Computational Linguistics arebetter supported by the arguments of thispaper.
AIM-158 CS-250 AD740127Ashok Chandra, Zohar Manna,Program Sclrelnas With Equality,13 pages, December 197 1.
We discuss the class of program schemasaugmented with equality tests, that is, tests ofequality between terms.
In the first part of the paper we illustrate the‘power’ of equality tests. It turns out that theclass of program schemas with equality ismore powerful than the ‘maximal’ classes ofschemas suggested by other investigators.
In the second part of the paper, we discussthe decision problems of program schemaswith equality. It is shown, for example, thatwhile the decision problems normallyconsidered for schemas (such as halting,divergence, equivalence, isomorphism andfreedom) are decidable for Ianov schemas.They all become undecidable if generalequality tests are added. We suggest,however, limited equality tests which can beadded to certain subclasses of programschemas while preserving their solvableproperties.
AIM- 159 CS-253Jerome A. Feldman, Paul C. Shields,Total Complexity and Iufereuce of BestPrograms,40 pages, April 1972.
A. 1. MEMO ABSTRACTS 105
Axioms for a total complexity measure forabstract programs are presented. Essentially,they require that total complexity be anunbounded increasing function of the Blumtime and size measures. Algorithms forfinding the best program on a finite domainare presented, and their limiting beha.vior forinfinite domains described. For totalcomplexity, there are important senses inwhich a machine can find the best programfor a large class of functions.
:::AIM- 160 CS-255Jerome A. Feldman,Automatic Programming,20 pages, February 1972.
AD740140
The revival of interest in AutomaticProgramming is considered. The research isdivided into direct efforts and theoreticaldevelopments and the successes and prospectsof each are described.
AIM-161 CS-264 AD741 189Yorick Wilks,Artificial Intelligence approach to MachilreTranslation,44 pages, February 1972.
The paper describes a system of semanticanalysis and generation, programmed InLISP 1.5 and designed to pass fromparagraph length input in English to Frenchvia an interlingual representation. A wideclass of English input forms will be covered,but the vocabulary will initially be restrictedto one of a few hundred words. With- thissubset working, and during the current year(197 l-72). it is also hoped to map theinterlingual representation onto somepredicate calculus notation so as to makepossible the answering of very simplequestions -about the translated matter. Thespecification of the translation system itself iscomplete, and its main points are:
i) It translates phrase by phrase--withfacilities for reordering phrases and
establishing essential semantic connectivitiesbetween them--by mapping complex semanticstuctures of “message” onto each phrase.These constitute the inter-lingualrepresentation to be translated. Thismatching is done without the explicit use ofa conventional syntax analysis, by taking asthe appropriate matched structure the ‘mostdense’ of the alternative structures derived.This method has been found highlysuccessful in earlier versions of this analysissystem.
ii) The French output strings are generatedwithout the explicit use of a generativegrammar. That is done by means ofstereotypes: strings of French words, andfunctions e’valuatmg to French words, whichare attached to English word senses in thedictionary and built into the interlingualrepresentation by the analysis routines. Thegeneration program thus receives aninterlingual representation that alreadycontains both French output and implicitprocedures for assembling the output, sincethe stereotypes are in effect recursiveprocedures specifying the content andproduction of the output word strings. Thusthe generation program at no time consults aword dictionary or inventory of grammarrules.
It is claimed that the system of notation andtranslation described is a convenient one forexpressing and handling the items ofsemantic information that are essential to anyeffective MT system. I discuss in some detailthe semantic information needed to ensurethe correct choice of output prepositions inFrench; a vital matter inadequately treatedby virtually all previous formalisms andprojects.
<<AIM- 162 CS-265 AD744634Roger C. Schank, N. Goldman, C. J. Rieger,C. K. Riesbeck,Primitive Concepts Underlying Verbs ofThought,102 pages, April 1972.
106
In order to create conceptual structures thatwill uniquely and unambiguously representthe meaning of an utterance, it is necessary toestablish ‘primitive’ underlying actions andstates into which verbs can be mapped. Thispaper presents analyses of the most commonmental verbs in terms of such primitiveactions and states. In order to represent theway people spea.k about their mentalprocesses, it was necessary to add to the usualideas of memory structure the notion ofImmediate Memory. It is then argued thatthere are only three primitive mental ACTS.
AIM-163 CS-266Jean M. Cadiou,Recursive Definitions of Partial Functionsalrd Their Conlputations,Thesis: Ph.D. in Computer Science,160 pages, April 1972.
A formal syntactic and semantic model ispresented for ‘recursive definitions’ which aregeneralizations of those found in LISP, forexample. Such recursive definitions can havetwo classes of fixpoints, the strong fixpointsand the weak fixpoints, and also possess aclass of computed partial functions.
Relations between these classes are presented:fix points are shown to be extensions ofcomputed functions. More precisely, strongfixpoints are shown to be extensions ofcomputed functions when the computationsmay involve ‘call by name’ substitutions;weak fixpoints are shown to be extensions ofcomputed functions when the computationonly involve ‘call by value’ substitutions.The Church-Rosser property for recursivedefinitions with fixpoints also follows fromthese results.
Th’en conditions are given on the recursivedefinitions to ensure that they possess leastfix points (of both classes), and computationrules are given for computing these twofixpoints: the ‘full’ computation rule, whichleads to the least weak fixpoint. A general
A. I. MEMO AskTRACTS
class of computation rules, called ‘safeinnermost’, also lead to the latter fixpoint.The “leftmost innermost” rule is a specialcase of those, for the LISP recursivedefinitions.
AIM- 164 CS-2 72 AD742748Zohar Manna, Jean Vuillemin,Fixpoirlt Approach to the Theory ofComputation,29 pages, April 1972.
Following the fixpomt theory of Scott, wepropose to define the semantics of computerprograms in terms of the least fix points ofrecursive programs. This allows one not onlyto justify all existing verification techniques,but also to extend them to handle variousproperties of computer programs, includingcorrectness, termination and equivalence, in auniform manner.
<<AIM- 165 CS-280 AD74275 ID. A. Bochvar,Two Papers 011 Partial Predicate Calculus,50 pages, April 1972.
These papers, published in 1938 and 1943,contain the first treatment of a logic ofpartial predicates. Bochvar’s treatment is ofcurrent interest for two reasons. First, partialpredicate and function logic are importantfor mathematical theory of computationbecause functions defined by programs or byrecursion cannot be guaranteed to be total.Second, natural language may be betterapproximated by a logic in which somesentences may be undefined than by aconventional logic. Bochvar use of hissystem to avoid Russell’s paradox is ofinterest here, and in partial predicate logic itmay be possible to get more of anaxiomatitation of truth and knowledge thanin a conventional logic.
The papers translated are On a three-valuedlogical calculus and its application to theanalysis of contradictions, Recueil
A. I. MEMO ABSTRACTS 107
Mathematique, N. S. 4 (1938), pp. 287-308,and 011 the comistemy of a three-valuedlogical calculus, ibid. 12 (1943), pp. 353-369.
small C. We also describe an algorithm thatlies between thee two: it takes time n log(n)and space log(n).
We also print a review and a correction byAlonzo Church that appeared in the Journalof Symbolic Logic. The review was in Vol.4.2 (June 1939), p. 99, and the additionalcomment was in Vol. 5.3 (September 1940), p.119.
It is shown that several problems are closelyrelated to the linear recursion problem, forexample, the problem of reversing an inputtape given a finite automaton with severalone-way heads. By casting all these problemsinto canonical form, efficient solutions areobtained simultaneously for all.
AIM- 166 cs-28 1 AD-743598Lynn H. Qam, S. Liebes, R. B. Tucker, M.J. Hannah, B. G. Eross,Computer biteractive Picture Processing,40 pages, April 1972.
AIM- 168 CS-287Shigeru Jgarashi,
AD746146
This report describes work done in imageprocessing using an interactive computersystem. Techniques for image differencingare described and examples using imagesreturned from Mars by the Mariner Ninespacecraft are shown. Also described aretechniques for stereo image processing.Stereo processing for both conventionalcamera systems and the Viking 1975 Landercamera system is reviewed.
Admissibility of Fixed-point Ilrductiorl inFirst-order Logic of Typed Theories,Diskfile: FIXPNT.IGR[AIM,DOC]40 pages, May 1972.
First-order logic is extended so as to dealwith typed theories, especially that ofcontinuous functions with fixed-pointinduction formalized by D. Scott. Thetranslation of his forma.1 system, or the 1calculus-oriented system derived andimplemented by R. Milner, into this logicamounts to adding predicate calculus featuresto them.
AIM- 167 CS -282Ashok K. Chandra,
AD747254
Efficient Cotnpilatiorl of Linear RecursivePrograms,43 pages, June 1972.
We consider the class of linear recursiveprograms. A linear recursive program is aset of procedures where each procedure canmake at most one recursive call. Theconventional stack implementation ofI-ecu rsion requires time and space bothproportional to n, the depth of recursion. ItIS shown that in order to implement linearrecursion - so as to execute in time n onedoesn’t need space proportional to n : nC forsufficiently small c will do. It is also knownthat with constant space one can implementlinear recursion in time n2. We show thatone can do much better: nl+C for arbitrarily
In such a logic the fixed-point inductionaxioms are no longer valid, in general, sothat we characterize formulas for whichScott-type induction is applicable, in terms ofsyntax which can be checked by machinesautomatically.
AIM- 169 CS-288Robin Milner,Logic for Computable Fumtiom:Descriptioll of a Machine Implementation,Diskfile: LCFMAN.RCM[AIM,DOC],36 pages, May 1972.
This paper is primarily a user’s manual forLCF, a proof-checking program for a logic ofcomputable functions proposed by DanaScott in 1969, but unpublished by him. Weuse the name LCF also for the logic itself,
108 A. 1. MEMO AIkTRACTS
which is presented at the start of the paper.The proof-checking program is designed toallow the user interactively to generateformal proofs about computable functionsand functionals over a variety of domains,Including those of interest to the computerscientist -- for example, integers, lists andcomputer programs and their semantics. Theuser’s task is alleviated by two features: asubgoaling facility and a powerfulsimplification mechanism. Applicationsinclude proofs of program correctness and inparticular of compiler correctness; theseapplicatipns are not discussed herein, but areillustrated in the papers referenced in theIntroduction.
i n conceptual representation is given.Certain adverbs are shown to berepresentative of complex belief structures.These adverbs serve as pointers that explainwhere the sentence that they modify belongsin a belief structure.
:::AIM- 172 cs-299Sylvia Weber Russell,
AD 75280 1
Setnarltic Categories of Nolllinals forCorlceptual Depetidewy Analysis ofNatural Language,64 pages, July 1972.
AIM-170 CS-289 AD748607Yorick Wilks,L,akof’f on Linguistics aud Natural Logic,Diskfile: LAK0FF.Y AW[AIM,DOCl19 pages, June 1972.
A system for the semantic categorization ofconceptual objects (nominals) is provided.The system is intended to aid computerunderstanding of natural language. Specificimplementations for ‘noun-pairs’ andprepositional phrases are offered.
The paper examines and criticizes Lakoff’snotIons of a natural logic and of a generativesemantics described in terms of logic, I arguethat the relationship of these notions to logicas normally understood is unclear, but Isuggest, in the course of the paper, a numberof. possible interpretations of his thesis ofgenerative semantics. I argue further that onthese interpretations a mere notationalv3 I i3nt of Chomskyan theory. I argue, too,that Lakoff’s work may provide a service inthat it constitutes a reductio ad dmcrdum ofthe derivational paradigm of modernlltl~ulstics; and shows, inadvertently, thatr)!ll\’ a system with the ability to reconsiderIt: own inferences can do the job that LakoffSPIS up for linguistic enquirey -- that is tosay, only an ‘artificial intelligence’ system.
::tAIM- 173 cs- 305 AD755139Gerald Jacob Agin,Represeutatiorr arrd Description of CurvedObjects,Thesis: Ph.D. in Computer Science,134 pages, October 1972.
Three dimensional images, similar to depthmaps, are obtained with a triangulationsystem using a television camera, and adeflectable laser beam diverged into a planeby a cylindrical lens.
Complex objects ar: represented as structuresjoining parts called generalized cylinders.These primitives are formalized in a volumerepresentation by an arbitrary cross sectionvarying along a space curve axis. Severaltypes of joint structures are discussed.
I :4 I M - I 7 t cs-290Roger Schank,
AD746147
Adverbs and Belief,90 pages, June 1972.
The treatment of a certain class of adverbs
Experimental results are shown for thedescription (building of internal computermodels) of a handful of complex objects,beginning with laser range data from actualobjects. Our programs have generatedcomplete descriptions of rings, cones, andsnake-like objects, all of which may be
-
A. I. MEMO ABSTRACTS
described by a single primitive. Complexobjects, such as dolls, have been segmentedInto parts, most of which are well describedby programs which implement generalizedcylmder descriptions.
.::AIM- 174 cs-303 PB212827Francis Lockwood Morris,Correctness of Translations ofProgramming Languages -- an AlgebraicApproach,Thesis: Ph.D. in Computer Science,124 pages, August 1972.
Prograrnniing languages and their sets ofmeanings can be modelled by genera1operator algebras; semantic functions andcompiling functions by homomorphisms ofoperator algebras. A restricted class ofindividual programs, machines, andcomputations can be modelled in a uniformmanner by binary relational algebras. Arestricted class of individual manner bybinary relational algebras. These twoappJications of algebra to computing arecompatible: the semantic function providedbY interpreting (‘running’) one binaryrelational algebra on another is ahomomorphism on an operator algebrawhose elements are binary relationalalgebras.
Using these mathematical tools, proofs can beprovided systematically of the correctness ofcompilers for fragmentary programminglanguages each embodying a single language‘feature’. Exemplary proofs are given .forstatement sequences, arithmetic expressions,Boolean expressions, assignment statements,and while statement. Moreover, proofs ofthis sort can be combined to provide(synthetic) proofs for in principle, manydifferent c_omplete programming languages.One example of such a synthesis is given.
109
AIM-175 cs-307Hozumi Tanaka,Hadaward Transforin for Speech WaveAnalysis,Diskfile: HADAM.HT[AIM,DOC],34 pages, August 1972.
Two methods of speech wave analysis usingthe Hadamard transform are discussed. Thefirst method is a direct application of theHadamard transform for speech waves. Thereason this method yields poor results ISdiscussed. The second method is theapplication of the Hadamard transform to alog-magnitude frequency spectrum. After theapplication of the Fourier transform theHadamard transform is applied to detect apitch period or to get a smoothed spectrum.This method shows some positive aspects ofthe Hadamard transform for the analysis ofa speech wave with regard to the reductionof processing time required for smoothing,but at the cost of precision. A formanttracking program for voiced speech isimplemented by using this method and ‘anedge following technique used in scenean al ysis.
({AIM- 176 G-308 AD754109Jerome A. Feldman, J. R. Low, D. C.Swinehart, R. H. Taylor,Recent Developinents in SAIL -- anALGOL based Language for ArtificialIntelligence,22 pages, November 1972.
New features added to SAIL, an ALGOLbased languaged for the PDP- 10, arediscussed. The features include: procedurevariables; multiple processes; coroutines; alimited form of backtracking; an eventmechanism for inter-process communication;and matching procedures, a new way ofsearching the LEAP associative data base.
110 A. I. MEMO ABSTRACTS
AIM-177 cs-3 11Richard Paul,Modelling, Trajectory Calculatiou aiidServoing of a Computer Controlled Arm,Thesis: Ph.D. in Computer Science,89 pages, November 1972.
The problem of computer contra1 of an arm_ ._IS divided into four parts: modelling,trajectory calculation, servoing and control.
In modelling we use a symbolic data structureto represent objects in the environment. Theprogram considers how the hand may bepositioned to grasp these objects and planshow to turn and position them in order tomake various moves. An arm model is usedto calculate the configuration-dependentdynamic properties of the arm before it ismoved.
The arm is moved along time-coordinatedspace trajectories in which velocity andacceleration are controlled. Trajectories arecalculated for motions along defined spacecurves, as in turning a crank; in suchtrajectories various joints must be free due toexternal motion constraints.
The arm is servocd by a small computer. Noanalog servo is used. The servo iscompensated for gravity loading and forconfiguration-dependent dynamic propertiesof the arm.
In order to control the arm, a planningprogram interprets symbolic arm cbntrolInstructions and generates a plan consistingof arm motions and hand actions.
The move planning program has workedsuccessfully in the manipulation of planefaced objects. Complex motions, such aslocating a bold and screwing a nut onto it,have also been performed.
({AIM- 178 CS-3 12 AD754 108Aharon Gill,Visual Feedback and Related Problems inComputer Controlled Hand eyeCoordination,Thesis: Ph.D. in Electrical Engineering,130 pages, October 1972.
A set of programs for precise manipulationof simple planar bounded objects, by meansof visual feedback, was developed for use inthe Stanford hand-eye system. The systemincludes a six degrees of freedom computercontrolled manipulator (arm and hand) anda fully instrumented computer televisioncamera.
The image of the hand and manipulatedobjects is acquired by the computer throughthe camera. The stored image is analyzedusing a corner and line finding programdeveloped for this purpose. The analysis issimplified by using all the informationavailable about the objects and the hand,and previously measured coordination errors.Simple touch and force sensing by the armhelp the determination of three dimensionalpositions from one view.
The utility of the information used tosimplify the scene analysis depends on theaccuracy of the geometrical models of thecamera and arm. A set of calibrationupdating techniques and programs wasdeveloped to maintain the accuracy of thecamera model relative to the arm model.
The precision obtained is better than .l inch.It is limited by the resolution of the imagingsystem and of the arm position measuringsystem.
<<AIM- 179 CS-320Bruce G. Baumgart,Winged Edge Polyhedron Representation,46 pages, October 1972.
A winged edge polyhedron representation is
A. I. MEMO ABSTRACTS 111
stated and a set of primitives that preserveEuler’s F-E+V=? equation are explained.Present use of this representation in ArtificialIntelligence for computer graphics and worldmodeling is illustrated and its intendedfuture application to computer vision isdescribed.
:::AIM- 180 CS-32 1 AD7597 12Ruzena Bajcsy,Computer Idelltificatiolr of TexturedVisual Scenes,Thesis: Ph.D. in Computer Science,156 pages, October 1972.
This work deals with computer analysis oftextured outdoor scenes involving grass, trees,water and clouds. Descriptions of texture areformalized from natural languagedescriptions; local descriptors are obtainedfrom the directional and non-directionalcomponents of the Fourier transform powerspectrum. Analytic expressions are obtainedfor orientation, contrast, size, spacing, and inperiodic cases, the locations of textureelements. These local descriptors are definedover windows of various sizes; the choice ofsjzes is made by a simple higher-levelprogram.
The process of region growing is representedby a sheaf-theoretical model which formalizesthe operation of pasting local structure (overa window) into global structure (over aregion). Programs were implemented whichform regions of similar color and similartexture with respect to the local descriptors.
An interpretation is made of texture gradientas distance gradient in space. A simpleworld model is described. An interpretationof texture regions and texture gradient ismade witb a simulated correspondence withthe world model. We find that a problem-solving approach, involving hypothesis-verification, more satisfactory than an earlierpattern recognition effort (Bajcsy 1970) andmore crucial to work with complex scenes
than in scenes of polyhedra. Geometric cluesfrom relative sizes, texture gradients, andinterposition are important in interpretation.
::eAIM-181 CS-325Bruce G. Buchanan,Review of Hubert Dreyfus’ ‘WhatComputers Can’t Do’: a Critique ofArtificial Reason,14 pages, November 1972.
The recent book “What Computers Can’tDo” by Hubert Dreyfus is an attack onartificial intelligence research. This reviewtakes the position that the philosophicalcontent of the book is interesting, but thatthe attack on artificial intelligence is not wellreasoned.
({AIM- 182 CS-326 AD754107Kenneth Mark Colby and Franklin DennisHilf,Call Expert Judges, using Trarlscripts ofTeletyped Psychiatric Illterviews,Distinguish Humarr Paranoid Patientsfrom a Computer Simulation of ParanoidProcesses?,10 pages, December, 1972.
Expert judges (psychiatrists and computerscientists) could not correctly distinguish asimulation model of paranoid processes fromactual paranoid patients.
::tAIM- 183 cs-344 AD759716Roger C. Schank,The Fourteen Primitive Actiom alld theirIuferelices,70 pages, March 1973.
In order to represent the conceptualinformation underlying a natural languagesentence, a conceptual structure has beenestablished that uses the basic actor-action-object framework. It was the intent tha tthese structures have only one representationfor one meaning, regardless of the semanticform of the sentence being represented.
112 A. I. MEMO ABSTRACTS
Actions were reduced to their basic parts soas to effect this. It was found that onlyfourteen basic actions were needed asbuilding blocks by which all verbs can berepresented. Each of these actions has a setof actions or states which can be inferredwhen they are present.
sAIM - 186 CS-332Robin Milner,Models of LCF,17 pages, January 1973.
AD758645
::!AIM-184 cs-330M alcolm Newey,
AD758651
Axiotns and Theoretns for Integers, Listsand Finite Sets in LCF,53 pages, January 1973.
LCF (Logic for Computable Functions) isbeing promoted as a formal languagesuitable for the discussion of variousproblems in the Mathematical Theory ofComputation (MTC). To this end, severalexamples of MTC problems have beenformalised and proofs have been exhibitedusmg the LCF proof-checker. However, inthese examples, there has been a certainamount of ad-hoc-ery in the proofs; namelyniany mathematical theorems have beenassumed without proof and noaxiomatisation of the mathematical domainsinvolved was given. This paper describes asuitable mathematical environment for futureLCF experiments and its axiomatic basis.The environment developed deemedappropriate for such experiments, consists ofa large body of theorems from the areas ofinteger arithmetic, list manipulation andfinite set theory.
LCF is a deductive system for computablefunctions proposed by D. Scott in 1969 in anunpublished memorandum. The purpose ofthe present paper is to demonstrate thesoundness of the system with respect tocertain models, which are partially ordereddomains of continuous functions. Thisdemonstration was supplied by Scott in hismemorandum; the present paper is merelyintended to make this work more accessible.
({AIM- 187 cs-33 1 AD757364George E. Collins,The Cotnputitlg Titne of the EuclideanAlgorithm,17 pages, January 1973.
The maximum, minimum and averagecomputing times of the classical Euclideanalgorithm for the greatest common divisor oftwo integers are der ived, to wi thincodominance, as functions of the lengths ofthe two inputs and the output.
(<AIM- 188 CS-336 AD 758646Ashok K. Chandra,On the Properties and Applications ofProgratn Schettras,Thesis: Ph.D. in Computer Science,23 1 pages, March 1973.
:::AIM- 185 cs-333 AD’157367Ashok K. Chandra, Zohar Manna,011 the Power of Progratntnitlg Features,29 pages, January 1973.
We consider the power of severalprograniming features such as counters,pushdown stacks, queues, arrays, recursionand equality. In this study program schemasare used as the model for computation. Therelations between the powers of these featuresis completely described by a comparisondiagram.
The interesting questions one can ask aboutprogram schemas include questions about the‘power’ of classes of schemas and theirdecision problems viz. halting divergence,equivalence, etc. We first consider thepowers of schemas with various features:recursion, equality tests, and several datastructures such as pushdown stacks, lists,queues and arrays. We then consider thedecision problems for schemas with equalityand with commutative and invertible
‘functions. Finally a generalized class of
A. I. MEMO ABSTRACTS
:chemas is described in an attempt to unifyhe various classes of uninterpreted andxmi-interpreted schemas and schemas withqpeclal data structures.
,sAIM-IS9 cs-337J~nles Gips, George Stiny,Aest lret its Systrn&
PB218682
22 pages, January 1973.
The formal structure of aesthetics systems isdefined. Aesthetics systems provide for thecwiti311 tasks of interpretation andt\ ,-1luarion , in aesthetic analysis.Ii olmogorov’s formulation of informationtheory is applicable. An aesthetics system fora class of non-representational, geometricpaintings and its application to three actualpaintings is described in the Appendix.
AIM-190 cs- 340h/l a lcolm Newey,
AD759714
Notes on a Problem IuvolviugPeriiiuta~ions as Sequences,20 pages, March 1973.
The problem (attributed to R. M. Karp byKnuth ) is to describe the sequences ofmlnimum length which contain, assubsequences, all the permutattons of analphabet of n symbols. This paper catalogssome of the easy observations on the problemand proves that the minimum lengths forn=5, n=6 and n=7 are 19, 28, and 39respectively. Also presented is a constructionwhich yields (for n>2) many appropriatesequences of length n -2n+42 so giving anupper bound on length of minimum stringswhich matches exactly all known values.
AIM-191 cs-34 1 AD 764272Shmuel M. Katz, Zohar Manna,A Heuristic-Approach to ProgramVerification,40 pages, March 1973.
We present various heuristic techniques foruse in proving the correctness of computer
113
programs. The techniques are designed toobtain automatically the “inductiveassertions” attached to the loops of theprogram which previously required human“understanding” of the program’s approaches:one in which we obtain the inductiveassertion by analyzing predicates which areknown to be true at the entrances and exitsof the loop (top-down approach), andanother in which we generate the inductiveassertion directly from the statements of theloop (bottom-up approach).
AIM-192 cs-345George E. Collins, Ellis Horowitz,The Minimum Root Separation of aPoly~io~ii i al,13 pages, April 1973.
The minimum root separation of a complexpolynomial A is defined as the minimum ofthe distances between distinct roots of A. Forpolynomials with Gaussian integercoefficients and no multiple roots, three lowerbounds are derived for the root separation.In each case the bound IS a function of thedegree, n, of A and the sum, d, of theabsolute values of the coefficients of A. Thenotion of a semi-norm for a commutativering is defined, and it is shown how anysemi-norm can be extended to polynomialrings and matrix rings,’ obtaining a verygeneral analogue of Hadamard’s determinanttheorem.
AIM-193 CS- 346Kenneth Mark Colby,
AD759717
The Rationale for Computer BasedTreatmerit of Language Difficulties inNonspeaking Autistic Children,Diskfile: AUTISM.KMC[AIM,DOC],13 pages, March 1973.
The principles underlying a computer-basedtreatment method for language acquisition innonspeaking autistic children are described.The main principle involves encouragementof exploratory learning with minimum adultinterference.
114 A. I. MEMO A’BSTRACTS
AIM-194 cs-347 PB22 1170/lKenneth Mark Colby, Franklin Dennis Hilf,Mult idimeusiotlal Allalysis irt Evaluatillg aSimulation of Paralroid Thought,10 pages, May 1973.
1 he limitations of Turing’s Test as anevaluation procedure are reviewed. Morevaluable are tests which ask expert judges tomake ratings along multiple dimensionsessential to the model. In this way themodel’s weaknesses become clarified and themodel builder learns where the model mustbe improved.
AIM-195 CS-356 PB222 164David Canfield Smith, Horace J. Enea,MLfSP2,Diskfile: MLISP2DAV[AIM,DOCl,LI 1 pages, May 1973.
MLISP2 is a high-level programminglanguage based on LISP. Features:
1. The notation of MLISP.12. Extensibility -- the ability to extend the
language and to define new languages.3. Pattern matching -- the ability to match
input against context free or sensitivepatterns.
4. Backtracking -- the ability to set decisionpoints, manipulate contexts andbacktrack.
,,AIM-196 cs-357 AD762471Nell M. Goldman, Christopher K. Riesbeck,A Conceptually Based SerrtenceParaphraser,Diskfile: MARGIE.NMG[AIM,DOCl,88 pages, May 1973.
This report describes a system of programswhich perform natural language processingbased -on an underlying language free(conceptual) representation of meaning. Thissystem is used to produce sentenceparaphrases which demonstrate a form ofunderstanding with respect to a givencontext. Parttcular emphasis has been placed
on the major subtasks of language analysis(mapping natural language into conceptualstructures) and language generation(mapping conceptual structures into naturallanguage), and on the interaction betweenthese processes and a conceptual memorymodel. .
({AIM- 197 CS-358 AD762470Roger C. Schank, Charles J. Rieger III,Inference alld the Computer Ullderstallditlgof Natural Language,63 psges, May 1973.
The notion of computer understanding ofnatural language is examined relative toinference mechanisms designed to function ina language-free deep conceptual base(Conceptual Dependency). The conceptualanalysis of a natural language seentence intothis conceptual base, and the nature of thememory which stores and operates uponthese conceptual structures are describedfrom both theoretical and practicalstandpoints. The various types of inferenceswhich can be made during and after theconceptual analysis of a sentence are defined,and a functioning program which performsthese inference tasks is described. Actualcomputer output is included.
AIM-198 CS-364 AD76361 1Ravindra B. Thosar,Estimation of Probability Detwity usirlgSigllature Tables for Applicatioll toPatter11 Recognition,37 pages, May 1973.
Signature table training method consists ofcumulative evaluation of a function (such asa probability density) at pre-assigned co-ordinate values of input parameters to thetable. The training is conditional: based ona binary valued ‘learning’ input to a tablewhich is compared to the label attached toeach training sample. Interpretation of anunknown sample vector is then equivalent ofa table lookup, i.e. extraction of the function
Ii -
A. I. MEMO ABSTRACTS 115
value stored at the proper co-ordinates. Sucha technique is very useful when a largenumber of samples must be interpreted as inthe case of samples must be interpreted as inthe case of speech recognition and the timerequired for the trainng as well as for therecognition is at a premium. However, thismethod is limited by prhibitive storagerequirements, even for a moderate number ofpara meters, when their relative independencecannot be assumed. This report investigatesthe conditions under which the higherdimensional probability density function canbe decomposed so that the density estimate isobtained by a hierarchy of signature tableswith consequent reduction in the storagereq u i remen t. Practical utility of thetheoretical results obtained in the report isdemonstrated by a vowel recognitionexperiment.
Defining the semantics of programminglanguages by axioms and rules of inferenceyields a deduction system within whichproofs may be given that programs satisfyspecifications. The deduction system hereinis shown to be consistent and also deductivecomplete with respect to Hoare’s sustem. Asubgoaler for the deductive system isdescribed whose input is a significant subsetof Pascal programs plus inductive assertions.The output is a set of verification conditionsor lemmas to be proved. Several non-trivialarithmetic and sorting programs have beenshown to satisfy specifications by using aninteractlve theorem prover to automaticallygenerate prrofs of the ved%-?tion conditions.Additional components for a more powerfulverficiatlon system are under construction.
:oAIM -20 1 cs-“f%Cunnar Rutger Grape,
AD763673
AIM-199 cs-398Bruce G. Baumgart,
AD771300
Image Contouring and Comparing,52 pages, (in preparation).
Model Based (Intermediate Level) ComputerVision,Thesis: Ph.D. in Computer Science,256 pages, May 1973.
A contour image representation is stated andan algorithm for converting a set of digitaltelevision images into this representation isexplained. The algorithm consists of fivesteps: digital image thresholding, binaryimage contouring, polygon nesting, polygonsmoothing, and polygon comparing. Animplementation of the algorithm is the mainroutine of a program called CRE; auxiliaryroutines provide cart and turn table control,TV camera input, image display, and Xeroxprinter output. A ser&dip application ofCRE to type font construction is explained.Details about the intended application ofCRE to the perception of physical objects~111 appear in sequels to this paper.
AIM-200 - CS-365Shlgeru Igarashi, David C. Luckham, RalphL. London,Automatic Program Verificatiolr I: LogicalBasis and its Implementation,50 pages, May 1973.
A system for computer vision is presented,which is based o n two-dimension alprototypes, and which uses a hierarchy offeatures for mapping purposes. Morespecifically, we are dealing with scenescomposed of planar faced, convex objects.Extensions to the general planar faced caseare discussed. The visual input is providedby a TV-camera, and the problem is to -interpret that input by computer, as aprojection of a three-dimensional scene. Thedigitized picture is first scanned forsignificant intensity gradients (called edges),which are likely to appear at region- andobject junctions. The two-dimensional scene-representation given by the totality of suchintensity discontinuitles (that word usedsomewhat inexactly) is referred to in thesequel as the ‘edge-drawing’, and constitutesthe input to the vision system presented here.
The system proposed and demonstrated in
-
116 A. I. MEMO ABSTRACTS
this paper utilizes perspectively consistenttwo-dimensional models (prototypes) of viewsof three-dimensional objects, andinterpretations of scene-representations arebased on the establish- ment of mappingrelatlonships from conglomerates of scene-elements (1’ine-constellations) to prototypetemplates. The prorotypes are learned by theprogram through analysis of - andgeneralization on - ideal instances. Thesystem works better than any sequential (orother! system presented so far. It should bewell suited to the context of a complete visionsystem e using depth, occlusion, supportrelations, etc. The general case of irregularlyshaped, planar faced objects, includingconcave ones, would necessitate such anextended context.
A I M -202 CS- 368 AD 764396Roger C. Schank, Yorick Wilks,The Goals of Linguistic Theory Revisited,44 pages, May 1973.
We examine the original goals of generativelmguistlc theory. We suggest that these goalswere well defined but misguided with respectto their avoidance the problem of modellingperformance. We developments such asGenerative Semantics, it is no longer clearthat the goals are clearly defined. We arguethat it is vital for linguistics to concern itselfwith the procedures that humans use inlanguage. We then introduce a number ofbasic human competencies, in the field oflanguage understanding, understanding incontext and the use of inferentialinformation, and argue that the modelling ofthese aspects of language understandingrequires procedures of a sort that cannot beeasily accomodated within the dominantparadigm. In particular, we a.rgue that theprocedures tha.t will be required in thesecases ought to be linguistic, and that thesimple-minded importation of techniques,and that the simple-minded importation oftechniques from logic may create a linguisticsIn which there cannot be procedures of therequired sort.
AIM-203 CS- 369 AD764274Roger C. Schank,The Development of Conceptual Structuresin Children,31 pages, May 1973.
Previous papers by the author havehypothesized that is is possible to representthe meaning of natural language sentencesusing a framework which has only fourteenprimitive ACTS. This paper addresses theproblem of when and how these ACTS mightbe learned by children. The speech of achild of age 2 is examined for possibleknpwledge of the primitive ACTS as well asthe conceptual relations underlying language.It is shown that there is evidence that theconceptual structures underlying language areprobably complete by age 2. Next a child isstudied from birth to age 1. The emergenceof the primitive ACTS and the conceptualrelations is traced. The hypothesis is madethat the structures that underlie and arenecessary for language are present by age 1.
SAIL is a high-level programming languagefor the PDP-10 computer. It includes anextended ALGOL 60 compiler and acompanion set of execution-time routines. Inaddition to ALGOL, the language features:(1) flexible linking to hand-coded machinelanguage algorithms, (2) complete access tothe PDP-10 I/O facilities, (3) a completesystem of compile-time arithmetic and logicas well as a flexible macro system (4) usermodifiable error handling, (5) backtracking,and (6) interrupt facilities. Furthermore, asubset of the SAIL language, called LEAP,provides facilities for (I) sets and lists, (2) anassociative data structure, (3) independentprocesses, and (4) procedure variables. TheLEAP subset of SAIL is an extension of the
A. I. MEMO ABSTRACTS 117
LEAP language, which was designed by J.Feldman and P. Rovner, and implementedon Lincoln Laboratory’s TX-2 (see [Feldman& Rovner]). The extensions to LEAP arepartially described in ‘Recent Developmentsin SAIL’ (see [Feldman]).
[TG] and generative semantics [GS], whichseeks to determine the well-formedness, orotherwise, of sentences.
This manual describes the SAIL languageand the execution-time routines for thetypical SAIL user: a non-novice programmerwith some knowledge of ALGOL. It liessomewhere between being a tutorial and areference manual.
AIM-205 cs-370N. S. Sridharan, et al,
AD764288
A Heuristic Program to Discover Synthesesfor Conlplex Organic Molecules,30 pages, June 1973.
I outline a system of preference semanticsthat does this: for each phrase or clause of acomplex sentence, the system builds up anetwork of lexical trees with the aid ofstructured items called templates and, at thenext level, it structures those networks withhigher level items called paraplates andcommon-sense inference rules. At each stagethe system directs itself towards the correctnetwork by always opting for the most‘semantically dense’ one it can construct. Isuggest- that this opting for the ‘greatestsemantic density’ can be seen as aninterpretation o f 300s’ ‘Semantic AxiomNumber 1’.
Organic Chemical Synthesis is found to be asuitable program for developing machineintelligence. A previous paper described theobjective and global characteristics of theproject. The present article aims to describethe program organization as a heuristicsearch, the design of the Problem SolvingTree and the search procedures inconsiderable detail. Examples of synthesesdiscovered and the problem solving treedeveloped are given. The programs arewritten mostly in PLl(F) applicable to anIBM 360/67 and the timings (batch mode)indicate that we have fast and efficientpractical systems.’
I argue that the analysis of quite simpleexamples requires the use of inductive rulesof inference which cannot, theoreticallycannot, be accomodated within thederivational paradigm. I contrast t h i sderivation al pa rad igm o f languageprocessing with the artificial intelligence [AI]paradigm.
AIM-207 CS-378 AD767333James Anderson Moorer,The ‘Optimum-comb’ Method of PitchPeriod Allalysis in Speech,25 pages, June 1967.
AIM-206 cs-377 - AD 764652Yorick Wilks,Preference Seniantics,20 pages, July 1973.
A new method of tracking the fundamentalfrequency of voiced speech is descirbed. Themethod is shown to be of similar accuracy asthe Cepstrum technique. Since the methodinvolves only addition, no multiplication, it isshown to be faster than the SIFT algorithm.
Preference semantics [PSI is a set of formalprocedures -for representing the meaningstructure of natural language, with a view toembodying that structure within a systemthat can be said to understand, rather thanwithin what I would call the ‘derivationalparadigm’, of transformational grammar
AIM-208 cs-379 AD767334James Anderson Moorer,The Heterodyne Method of Analysis ofTransient Waveforms,25 pages, June 1973.
118
A method of analysis of transient waveformsis discussed. Its properties and limitationsare presented m the context of musical tones.The method is shown to be useful when therIsetimes of the partials of the tone are nottoo short. An extention to inharmonicpartials and polyphonic musical sound isdiscussed.
A I hl -209 cs-380 AD767695Yora m Y akimovsky,Scene Analysis using a Semautic Base forRegion Growing,Thesis:. Ph.D. in Computer Science,120 pages, July 1973.
The problem of breaking an image intomPanlngfL11 regions is considered. Aprobabilistic semantic basis is effectivelyintegrated with the segmentation process,providing various decision criteria. Learningiacillties are provided for generatingi ti teractively the probabilistic bases. Aprogramming system which is based on theseideas and its successful application to twoproblem domarns are described.
NM-210 CS- 382 AD 767335Zohar Manna, Amir Pnueli,Axiomatic Approach to Total Correctnessof Programs,25 pages, July 1973.
We present here an axiomatic approachwhich enables one to prove by formalmethods that his progra.m is ‘totally correct’(I.e., it terminates and is logically cdrrect --does what it is supposed to do). Theapproach is similar to Hoare’s approach forproving that a program is ‘partially correct’\I e., that whenever it terminates it producescorrect results). O u r extension to Hoare’smethod- lies in the possibility of provingcorxtness and termination at once, and inthe enlarged scope of properties that can beproved by it.
![Mart´ın Abadi · 2005-09-16 · Systems 8, 1 (February 1990), 18–36. 50. “Temporal Logic Programming” [Mart´ın Abadi and Zohar Manna] Journal of Symbolic Computation 8,](https://static.documents.pub/doc/80x56/5f0b51ef7e708231d42feee8/martn-2005-09-16-systems-8-1-february-1990-18a36-50-aoetemporal-logic.jpg)
