A-A105 876 NAVAL OCEAN SYSTEMS CENTER SAN DIEGO

CA -/ /

C', RELEVANCE OF SELECTED ARTIFICIAL INTELLIGENCE SYSTEMS. BRIEF--ETC(UI.JL81 R .J I ECHTEL

UNCLASSIFIED NOSC/TR-706 NL

) LEVEL' *

S ADA 105876 .

Technical Report 706

C2 RELEVANCE OF SELECTED -ARTIFICIAL INTELLIGENCE SYSTEMS

Brief examination of several existing Al systems foruse in future command and control applications .

RJ Bechtel

DTIC 7 July 1981II EECTE 3 Interim Report for Period I October 1980-30 April 1981

,, OCT 2 0 1981

SPrepared forNaval Electronic Systems Command

Code 613

Approved for public release; distribution unlimited

AFNAVAL OCEAN SYSTEMS CENTER• SAN DIEGO, CALIFORNIA 92152

t 0 20

Bifea iaionI of seea exs i Alsytesous-n uur omad n coto apliaton

NAVAL OCEAN SYSTEMS CENTER, SAN DIEGO, CA 92152

AN ACTIVITY OF THE NAVAL MATERIAL COMMAND

SL GUILLE, CAPT, USN HL BLOODCommander Technical Director

ADMINISTRATIVE INFORMATION

Work was performed under Program Element 62721 N, Project F211241, Task AreaXF2 1241100 (NOSC 824-CC96), by a member of the C2 Information Processing Branch(Code 8242) for Naval Electronic Systems Command, Code 613. This report covers workfrom I October 1980 to 30 April 1981 and was approved for publication 7 July 1981.

Released by Under authority ofRC Kolb, Head JH Maynard, HeadTactical Command and Command Control -Electronic

Control Division Warfare Systems andTechnology Department

C

S

UNCLASSIFIED ____

SECURITY CLASSIFICATION OF THIS PAGE (Whien Dae. Entered)

REPOT DCUMNTATON AGEREAD INSTRUCTIONS______ REPORT___DOCUMENTATION ____PAGE_ BEFORECOMPLETINGFORM

1. REORT NMBER2. GOVT ACCESSION No. 3. RECIPIENT'S CATALOG NUMBER

E ~ 1t~ & ERIOD COVERED

:2TAELEVANCE OF SELECTED ARTIFICIAL INTELLIGENCE Interim-STEMS I Octlw 80- 30 ApitI% I

Brief ,xamination of VveraI, ;xisting Al systems for use in future 6. PERFORMING ORG. REPORT NU"ER

comn dand control avlications-T- Ra TH s IV II. OON I GRT0 ANT NUM6J'Rrp)

RJ Bechtel pT /~CNRC

9. PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROjECT. TASK

AREA II WORK UNIT NUMBERS

Naval Ocean Systems Center621NF14,

16.~~~ DISTRBUTIO STTMN Ao G~ ReporI

Waprnoved Dor public rees;diti4inniie

17. DISTRIBUTION STATEMENT (of thisbte needI lc 0 f ~frn t. Report)

IS. SUPPLEMENTARY NOTES

19. KEY WORDS (Conlirnuo on reverse *#do If neco. eay and identify. by block numb er)

Artificial intelligence Inference systemsKnowledge representation systems Natural language processingKnowledge presentation systems Planning and problem solving

20 ASYACT1(Cont)is m ine~aA#.l UedeWaty and identify by block nm.ber)

(A'very brief examination of several existing artificial intelligence systems which have been judged to beworth further investigation for use in future command and control applications. 01fis-ttsewntrepresentsthe current state of an ongoing process which began with a much longer list of artificial intelligence systems.Each System on the long list was given a very quick evaluation on stcales measuring availability, applicability,and maturity. Within each broad category of artificial intelligence effort (%gnatural language, inference)the highest ranking systems were selected for further examination. This is the result of that examination.

D0 1 'JAR7 1473k EDITION Of NOV 66aIS OU5IIOLSTE UNCLASSIFIEDS IN 0102- LIF 014- 6601 SECUITY CLASSIFICATION Ofr THIS PACE (t'Uen Dete Enteted)

OBJECTIVE

Survey and evaluate existing artificial intelligence systemsfor applicability to command and control. Include assessments ofavailability and maturity of techniques.

RESULTS

Brief summaries of sixteen artificial intelligence systemsare presented in five broad categories. Evaluations of thesystems are given, with a description of the evaluation criteria.A list of nine additional systems is included to direct possiblecontinuation of this effort.

RECOMMENDAT IONS

Use the evaluations in this report to direct research intothe application of artificial intelligence techniques to commandand control, and to focus transition efforts. If additionalsystem evaluations are done, examine the suggested additionalsystems first.

'1

Accession For

WTIS CFA&IPTIC TAM H]Unanrou f' Od El

Ava l h'A'Illtry Codes

Av&-I nrd/or

DiAt

AmL

S ,P=..

CONTENTS

1 Introduction . . . page 1

1.1 Availability . . . 11.2 Applicability . . . 11.3 Maturity . . . 2

2 Knowledge Representation Systems . . . 2

2.1 UNITS . . . 32.2 KL-ONE . . . 42.3 KNOBS . . . 5

3 Knowledge Presentation Systems. . . 6

3.1 AIPS . . . 73.2 SDMS. . . 7

4 Inference Systems . . . 8

4.1 E-MYCIN . . . 94.2 ROSIE . . . 10

4.3 STAMMER . . . 10

5 Natural Language Processing . . . 11

5.1 ATNs . . . 125.2 Yale work . . . 13

6 Planning and Problem Solving . . . 14

6.1 STRIPS, ABSTRIPS, and NOAH . . . 146.2 TALESPIN . . . 166.3 Other SRI Systems . . . 17

7 Original List and Preliminary Evaluations . . . 18

8 Other Systems Suggested for Inclusion . . . 21

9 Conclusions . . . 21

10 Recommendations . • . 21

ii

-~ _______

____

1 Introduction

This paper presents a very brief examination of severalexisting artificial intelligence systems which have been judgedto be worth further investigation for use in future command andcontrol applications. Rather than developing all support softwarefrom ground zero, it is desirable to make use of existingmaterial and avoid reinventing the wheel.

This document represents the current state of an ongoingprocess which began with a much longer list of artificial 4intelligence systems. Each system on the long list was given avery quick evaluation on scales measuring availability, -"

applicability, and maturity. Within each broad category ofartificial intelligence effort (e.g. natural language, inference)the highest ranking systems were selected for furtherexamination. This is the result of that examination. The originallist and evaluations are included as section 7 of this report.

This is a working document. It is anticipated that somesystems were missed and that some evaluations (particularlyavailability) may be incorrect. Comments, -ecommendations, andsuggested additions are welcomed.

1.1 Availability

High availat"iity ratings indicate that the system isreadily available in usable form, usually on a compatible host onthe ARPAnet. Evaluations were lowered for systems on incompatiblecomputers, systems which were old enough to probably havedisappeared, systems for which documentation or descriptions werelacking, and systems which would require substantial supportsystems (e.g. speech processing).

Unless otherwise noted, all systems examined run inINTERLISP-10 on a DEC-10 or DEC-20 that is a host on the ARPAnet.

1.2 Applicability

Aplicability was judged primarily by evidence of previouswork in areas related or similar to command and control. Often,such evidence was not available. When there was no evidence todraw upon, the underlying principles were relied upon forindicators of applicability. This approach resulted in ratingmost systems *possibly applicable,* which made applicabilityuseless as a fine-grained discriminator. However, some systems(like chess players) were clearly inapplicable.

I

1

1.3 Maturity

Systems whose working principles and theoretical limits arewell understood by most workers in the field are highly rated inmaturity. Systems understood only by their creators or still inearly design stages are rated low for maturity. Most systems fallsomewhere in between, with most workers understanding theapproach, a smaller set being conversant with the principles ofoperation, and a very few, if any, able to define the theoreticaland practical limitations of the system.

2 Knowledge Representation Systems

Intelligent systems or programs rely on two sources fortheir power. The first is well designed algorithms andheuristics, which specify the processes to be undertaken inpursuit of a task. The second is accurate, well organizedinformation about the task domain, which provides material forthe algorithms and heuristics to work with. The command andcontrol task domain has a wealth of information associated withit, but to date very little of this information has been madeavailable in a useful form for artificial intelligence programs.Organization of the information is as important as its capture,because poorly organized data will either (at best) slow thesystem or (at worst) be inaccessible.

Knowledge representation systems seek to provide frameworksfor organizing and storing information. Designers of differentsystems perceive different problem areas that need work, and thusdifferent systems do different things well. The goal must be toevaluate the strong and weak points of various representationsystems along with the requirements imposed by the informationpresent in the command and control task domain, so that anysystem selected will be capable of effective performance in thetask area, even if it is weak in other areas.

Command and control as a problem domain presents severalchallenges for knowledge representation systems. Two criticalfactors in C2 -- time and space -- were considered inadequatelyunderstood in a recent survey of workers in the field [1]. Aknowledge representation for C2 tasks must have some method forcoping with these problems as well as the problems of inheritance(e.g. making sure that all Kynda class platforms have the samemaximum speed capability) and synergy (e.g. 25 MIG fighters

[li SIGART Newsletter, Special Issue on KnowledgeRepresentation, edited by RJ Brachman and BC Smith. (70):1-138,February, 1980.

2

acting together under a unified command can inflict more damagethan 25 single fighter raids).

Since these issues must be faced, it is reasonable toconsider in greater depth those representation systems which haveaddressed some of these issues before considering those whichhave not. The systems considered here each include some form ofinheritance capability. Also, these representation systems appearto have sufficient flexibility to permit the embedding of othersystems for special-purpose representation, such as for time andspace.

2.1 UNITS

UNITS (21 is a package of LISP functions developed at 4Stanford for creating, deleting, accessing, updating, andotherwise manipulating knowledge representations organized aspartitioned semantic networks. In UNITS, the nodee of the networkare called units and the links among them are called slots. Units 1are connected in a generalization hierarchy which permits severalmodes of property inheritance. The package is available inINTERLISP on the ARPAnet.

Partitioned semantic networks have been proposed as a methodof extending the semantic network approach to permitquantification and other desirable features which are not obviousin most "standard" semantic network implementations. The UNITSpackage is oriented toward this framework for knowledgerepresentation and thus includes any shortcomings that theframework has. Specifically, the criticisms of Brachmanconcerning the "level" of the representation [3] are validremarks concerning UNITS.

On the other hand, UNITS has included some features beyondpartitioning which many semantic networks do not have. It ispossible to attach a procedure to a slot so that the procedurewill be executed when appropriate conditions are met -- forexample, when the slot is filled.

121 An Examination of a Frame-Structured Representation System,by M Stefik. In Proceedings of the Sixth International JointConference on Artificial Intelligence, pages 845-852. Tokyo,1979.

(31 On the Epistemological Status of Semantic Networks, by RIBrachman. BBN Report 3807, Bolt, Beranek, and Newman, April,1978.

3

hi- - - - , .. m l~

The inheritance hierarchy and defined inheritance modes alsoimpose some much-needed structure on an often nebulous area.

The most attractive feature of the UNITS system is thecompleteness of the support function library, which indicates acertain maturity. Not only are the "base* network manipulationfunctions provided, but there is also an editor which knows aboutslots and units and their structure and a package for displayingnetworks, which makes it much easier to tell what has been addedor changed.

In addition, the designers of UNITS have recognized thatmany applications will require quite large networks, often biggerthan can be readily maintained in core. To support such largenetworks, a "paging" facility has been provided so that usersneed not concern themselves directly with the management ofmemory space.

To date, there have been no efforts to represent navaldomains in UNITS, but the similarities between it and KL-ONE arepromising.

2.2 KL-ONE

KL-ONE [4] is a package of LISP functions developed at BBNfor creating and working with a particular form of associativenetwork representation called a strnctured inheritance npt. As inevery associative (semantic) network, the basic entities arenodes and links. KL-ONE shares with UNITS the idea of "typed"nodes, where nodes are required to be of certain limited classes.Also like UNITS, the type of a node constrains the links (slotsin UNITS) that can join it to other nodes. However, UNITS permitsthe user to define new slots that can also be used to link nodesand applies its type restrictions only on inheritance-relatedlinks. KL-ONE, by contrast, provides a base or primitive set oflinks along with the defined node types and does not permit thedefinition of new links. While this refusal to permit new linksmay impede flexibility, the claim is that links are always well-defined and never subject to the whims of users. KL-ONE is alsoavailable on the ARPAnet in INTERLISP.

KL-ONE was designed in an attempt to provide anepistemologically explicit representation system, with thedesigner's and user's assumptions open to observation in

(41 KLONE Reference Manual, by RJ Brachman, E Ciccarelli, NRGreenfeld, and MD Yonke. BBN Report 3848, Bolt, Beranek, andNewman, July, 1978.

4

representations built using the system. Two central aspects ofobjects which were felt to be essential to a successfulrepresentation paradigm were the inheritance of properties andthe description of structure. UNITS also addresses the problemof inheritance but provides no specific facilities for structuraldescription or specification.

KL-ONE gains particular appeal from its use as a foundationfor other systems, notably the AIPS system [5]. The AIPS work hasused KL-ONE to represent at least simple naval domaininformation. The fact that another application with apparentlyhigh relevance has found it useful is certainly encouraging.Moreover, in using KL-ONE, the AIPS developers have devised somevery useful support functions that make the creation andexamination of large networks much easier. Thus, KL-ONE isbeginning to approach the ease of user interface that is sopromising in UNITS.

2.3 KNOBS

KNOBS [6] is an integrated collection of artificialintelligence programs directed toward the development ofexperimental tactical air command and control. Developed atMITRE, KNOBS exists in INTERLISP on an ARPAnet host. While thereare several components to KNOBS, it is classified under knowledgerepresentation systems because that is where the majority of theeffort to date has been concentrated. Building on the frameworkof FRL, the KNOBS developers have translated the representationsystem into INTERLISP and have augmented it to permit interactiveframe instantiation. This insures that related items, such as thetype of aircraft chosen for a mission and their armament, areconsistent.

KNOBS is clearly oriented toward military applications. Forthis reason, there is much in it that is directly applicable tonaval command and control. It is also an integrated system,combining a powerful knowledge representation with a naturallanguage interface and some simple inferential capability.

(5] Application of Symbolic Processing to Command and Control, IFinal Report, by F Zdybel, MD Yonke, and NR Greenfeld. BBNReport 3849, Bolt, Beranek, and Newman, November, 1979.

[61 KNOBS: An Experimental Knowledge Based Tactical Air Mission IPlanning System and a Rule Based Aircraft IdentificationSimulation Facility by C Engelman, CH Berg, and M Bischoff. InProceedings of the Sixth International Joint Conference onArtificial Intelligence, pages 247-249. Tokyo, 1979.

5

Future plans include the addition of a rule-based inferencesystem. While the level of capability and integration isadmirable, KNOBS is still very much a system under developmentand as such is subject to at least minor change over the nearterm.

3 Knowledge Presentation Systems

Artificial intelligence programs, as noted earlier, usuallyrequire large amounts of information about the domain in whichthey operate. In addition to this domain knowledge, there isusually also a large amount of information within the domainwhich the program will process to perform its function. Forexample, a tactical situation assessment program's domainknowledge may include some formats (representations) for storinginformation about platforms, sensors, and sightings, while theinformation that it actually uses in doing assessment concernsreal platforms and sightings by sensors, organized as directed bythe representations.

Users should have access to the information within thedomain that is used by the program, both because it may be usefulin raw form and as a check on the program's operation. Managingthe presentation of such information is a complex task which hasnot been as well explored as the problems of informationacquisition.

The most widely used technique of knowledge presentation todate has been ordinary text. Occasionally the presentation isorganized in a question-answering form, but more commonly it isnot under user control at all. It has been especially difficultfor users to tailor the information presented to match theirneeds, concerns, and preferences. Presentation modes other thantext have also been extremely limited.

In command and control applications, flexibility to matchpresentation to needs is essential, as information requirementsvary with the situation. Different users may wish to vary thepresentation to emphasize points that each considers important,while a single user may wish to vary the presentation over timeto satisfy his current information requirements.

Concentration on readily reconfigurable graphics-basedpresentation systems would be particularly appropriate for C2,where the tool (graphics) can be well matched to the domain(spatial information). Another possible presentation mode thatcould demonstrate relevance would be synthesized speech output,to utilize a different sensory channel which may be more readilyavailable in some situations.

6AA-_~

3.1 AIPS

AIPS (Advanced Information Presentation System) [7] is anintelligent graphics interface designed at BBN especially tosupport command and control decision tasks. It utilizesrepresentations of knowledge about the domains of graphics,command and control, the display device, and the user's needs toprovide a highly interactive, individually tailored graphicdisplay subsystem for use by other systems in the command andcontrol domain. The underlying system for the various knowledgerepresentations is the representation system KL-ONE.

AIPS is available on the ARPAnet, but presumes theavailability of a particular bit-mapped graphics display devicefor successful operation. It does not use GL2 for graphiccommunication.

Graphics is one of the most useful modes of informationpresentation and one of the most difficult to manage. Thedifficulty in use arises from both the complexity of the graphicsdomain and the lack of standardization among output devices, bothof which can impose a heavy programming burden on those whochoose to incorporate graphic capability in their systems. AIPSholds out the promise of lifting the graphics interface burdenfrom the developers of other systems, while also making possiblegraphic presentation specialized for user needs. As mentionedbefore, some naval-domain-related work has been done in AIPS.

3.2 SDMS

The Spatial Data Management System (SDMS) [8], while notstrictly an artificial intelligence system, is an interestingapproach to information presentation, particularly of data whichwould usually be maintained in a traditional database system.SDMS was developed by Computer Corporation of America and isavailable as an experimental product. Instead of accessinginformation through a formal query language, SDMS presents theinformation graphically in a form which seems to encouragebrowsing and which requires less prior knowledge of the databasecontents and organization. Data are organized and retrieved by

[7] AIPS: An Information Presentation System for Decision Makersby NR Greenfeld and MD Yonke. BBN Report 4228, Bolt Beranek andNewman Inc., December, 1979.

(8] A Prototype Spatial Data Management System by CF Herot, RCarling, M Friedell, and D Kramlich. In SIGGRAPH 80 ConferenceProceedings, pages 63-70, 1980.

7

| |

positioning them in a Graphical Data Space which is viewedthrough a color raster scan display, permitting both surfacetraversals and zooms for greater detail.

SDMS is oriented toward the presentation of relativelystatic, slowly changing information of the type found in atraditional database. It permits an easily learned, highlyinteractive user interface to such static data but appears tohave no provisions for data addition and update, other thanthrough a somewhat laborious redefinition process. While somenaval domain data are certainly sufficiently static for this tobe a useful system (as shown by the demonstrations in thedescription), many more data are highly dynamic and so probablynot well suited by this particular system. In addition, thesystem is highly biased toward human use and requires aduplication of the internal database to permit access by bothprograms and the user.

4 Inference Systems

Inference is the process of drawing conclusions, of addinginformation to a knowledge (data) base on the basis ofinformation that is already there. Inference may be deductive orinductive, and a given system may permit both forms. Inferencesystems may operate in many different ways. One of the mostuseful forms is that of rule-based systems. Here, knowledge isstructured in rules which are applied to the facts to reachconclusions. The method of rule application forms the processbase for inference; while the rules are, in some sense, theknowledge-structuring base.

Most of the tasks in the command and control domain requiresome form of inference, of drawing conclusions from known facts.Again, tactical situation assessment is a good example. From(possibly limited) low-level information like sensor reports, itis necessary to reach conclusions about the existence andidentity of platforms in the vicinity.

An inference system for C2 tasks must be capable of dealingwith information common to the C2 domain. In practice, this willoften mean that the inference system must be adaptable enough towork in or with the knowledge representation framework chosen fora task. Flexibility with respect to knowledge representationtherefore becomes a major criterion in evaluating systems.

Another important aspect is the extensibility of theinference system. This is one area in which rule-based systemsare particularly attractive because, in general, new rules meannew inferences. Among rule-based systems, ease of rule additionand support of the addition function are important.

I8

.- --

4.1 E-MYCIN

The original MYCIN [91 was a rule-based system developed atStanford for diagnosing bacterial infections of the blood.E-MYCIN (101 is essentially the same system, but with the domain-specific information like the rules removed. E-MYCIN operates ina backward-chaining fashion -- given a goal, or statement to beconcluded, it applies rules in an attempt to find supportingevidence for the goal. This approach is also called goal-driveninference. It is available on the ARPAnet.

The representation available for rules in E-MYCIN isreasonably sophisticated, and there is a collection of functionswhich permit a form of reasoning under uncertainty. A relatedrule addition facility (called TEIRESIAS [11]) may also beavailable, but the work was done some time ago and may havedisappeared. Descriptions of TEIRESIAS are available. However,the available representation for facts is fairly sparse,especially in comparison to systems like UNITS and KL-ONE. Also,it is not clear that a synthesis of systems to overcome thisdeficiency is possible, and such an inability would impose strictlimits on the applicability of E-MYCIN.

Though several problem domains other than blood infectionshave been investigated using E-MYCIN, none of these involved anaval domain. Most applications have been in medicine.

[9] Computer-Baaed Medical Consultations! MYCIN, by EHShortliffe. American Elsevier, 1976.

[10] A Domain-Independent Production-Rule System forConsultation Programs, by W van Melle. In Proceedings of theSixth International Joint Conference on Artificial Intelligence,pages 923-925, 1979.

11] Knowledge Acquisition in Rule-Based Systems: Knowledgeabout Representation as a Basis for System Construction andMaintenance, by R Davis. In DA Waterman and F Hayes-Roth(editors), Pattern-Diraetad InferenCe ystem-, Academic Press,1978.

9

4.2 ROSIE

ROSIE [12] is another rule-based system which was developedat Rand as a successor to RITA [131, the Rand IntelligentTerminal Agent. ROSIE applies its rules in a forward-chainingmanner, going from facts to conclusions. This makes it data-driven rather than goal-driven. The underlying knowledgerepresentation is a common object-attribute-value structure,which has been used in many applications. ROSIE is available onthe ARPAnet. A recent major revision has resulted in a systemquite different from that available when this survey was started.

A design document exists for the original version. Itdetails a number of features which were not implemented in thatversion. Some of these features were added in the recentrevision. The structures available for facts appear richer thanthose of E-MYCIN, though they still fall short of the moresophisticated knowledge representation systems (e.g. inheritancemechanisms are not provided, though presumably the user could addthem). A synthesis of ROSIE and a more general knowledgerepresentation system might not be as difficult. NOSC personnelhave used ROSIE (both in its original incarnation and as revised)to implement rule-based inference systems in naval domains, soapplicability is high. However, these same implementations haverevealed some awkwardness forced on the user due, at least inpart, to the gaps among ROSIE design, implementation, anddocumentation as well as sparseness of knowledge representation.Past experience does indicate that some command and control tasksare well suited to a data-driven approach.

4.3 STAMMER

STAMMER [14] is a rule-based inference system developed atthe Naval Ocean Systems Center for use in tactical situationassessment. Like ROSIE, STAMMER is a forward-chaining system. The

[12] Design of a Rule-Oriented System for Implementing Expertise,by DA Waterman, RH Anderson, F Hayes-Roth, P Klahr, G Martins,and SJ Rosenschein. Rand Note N-1158-1-ARPA, The RandCorporation, May, 1979.

[13] RITA Reference Manual, by RH Anderson, M Gallegos, JJGillogly, RB Greenberg, and R Villanueva. Technical Report R-1808-ARPA, The Rand Corporation, 1977.

[141 STAMMER2 Production System for Tactical SituationAssessment, by DC McCall, PH Morris, DF Kibler, and RJ Bechtel.TD 298, Naval Ocean Systems Center, San Diego, CA, October, 1979.

10

knowledge representation used is a collection of independentassertions, with no structuring mechanism provided in therepresentation itself. The rule interpreter implements asuspension mechanism for improved time efficiency in repeatedrule applications. STAMMER is available on the ARPAnet.

The STAMMER system has been used in three experimentalapplications to date. The first of these, merchant detection,motivated the addition of graphic display support to the basicsystem. STAMMER has a moderately sophisticated user interface,with explanation and data base query capabilities among theactions available. Its applicability to tasks in the naval domainis proven, though it is somewhat biased toward the situationassessment task and is not an "empty" system. The lack offacilities for structuring the database is also a problem.

5 Natural Language Processing

The ability to use a natural language such as English tocommunicate with a computer has long been a goal of artificialintelligence researchers. A language understanding andgenerating capability could conceivably remove many obstaclesthat presently obstruct the human-machine interface. Naturallanguage processing, as examined here, is restricted to printedinput and output, with speech recognition and generation fallingunder a different heading.

Natural language processing (NLP) is relevant to command andcontrol in at least two ways. First, much of the present datathat needs to be captured for use by an automated C2 system arepassed in natural language form (e.g. as "free text" fields inmessages). NLP could assist in the capture of this information.Second, NLP could be used to improve the interface between theautomated system and its users, making the interface friendlierand more responsive.

One great problem with existing NLP systems is theirrestrictions on vocabulary and subject area. These restrictionsgrow from the need to provide domain knowledge to aid in languageunderstanding. For command and control, these restrictions are abenefit. The command and control domain is restricted, and it isreasonable to limit interactions to topics within the domain.This indicates that a limited vocabulary may well be sufficient,so that what would be problems in applying NLP to some otherdomain may be lessened in command and control.

11

5.1 ATNs

Recursive augmented transition network parsers (ATNs) havebeen studied and discussed in the literature for several years.The work that brought them to prominence was done by Woods andothers at BBN [15,161. These parsers work by performing statetransitions depending on the value of the word they arepresented. The possible transitions form a network, which may beentered recursively. Performing a transition may cause somefunction to be executed, possibly changing the value of one ofthe registers which augments the transition net. ATNs have thetheoretical power of Turing machines. ATNs are readilyaccessible. About the worst that could be expected would be atranslation from some other LISP dialect. The problem with usingan ATN is the writing of grammar and support routines, which willhave to be done regardless of the source of the parser itself.

ATNs are very well understood. They have been studied anddiscussed in the literature for several years. Further, many ATNsare semistandardized, making grammars written for one usuallyadaptable to another. Even when an existing ATN is not available,implementing one and providing a grammar is almost a cookbookproblem [171.

ATNs tend to rely heavily on syntactic features, since thoseare more readily codified than, say, semantics. In theory, ATNsare not biased toward any particular aspect of language, but inpractice most implementations are extended with semantic andpragmatic routines that lie outside the parser proper. Theflexibility of ATNs over problem domains has been welldemonstrated. However, providing a grammar and any extensions tonaval domains would be a user responsibility.

[15] Transition Network Grammars for Natural Language Analysis,by W Woods. Communications of the ACM 13:591-606, 1970.

[16] The LUNAR Sciences Natural Language Information System:Final Report, by WA Woods, RM Kaplan, and BL Nash-Webber. BBNReport 2378, Bolt Beranek and Newman Inc., 1972.

[171 Artificial Intaligence Programming, by E Charniak, CKRiesbeck, and DV McDermott. Lawrence Erlbaum Associates,Hillsdale, New Jersey, 1980.

12

• _ • __ _7- 7'r

5.2 Yale work

Over the past few years, a great deal of work by RogerSchank and his colleagues has come out of Yale in naturallanguage processing [18,19,20,21,22]. The central thesis ofmost of these efforts has been that language processing is ananticipatory activity, that language is understood by continuallymaking predictions and then testing those predictions against theactual input.

Also, this group has worked on providing conceptualframeworks for language constructs larger than single words orphrases to account for observed features of human language-understanding behavior. The Yale work exists primarily off theARPAnet in the UCILISP dialect, which would require transport andtranslation. Work using the same principles has been done orduplicated elsewhere (e.g. CMU) and so might be more availablethan the original Yale work.

Since the Yale approach to NLP is far different from that ofmost other workers, comparisons are very difficult. Only duringthe past year did a study of the differences between a Yale-

designed parser and an ATN finally describe the theoreticaldistinctions.

The Yale work is very interesting in its strong emphasis ondomain knowledge to guide understanding. However, this imposes alarge burden of domain analysis and encoding on anyone who

[181 Analyzing English Noun Groups for their Conceptual Content,by AV Gershman. Research Report 110, Department of ComputerScience, Yale University, May, 1977.

[19] Adaptive Understanding: Correcting erroneous inferences, byRH Granger. Research Report 171, Department of Computer Science,Yale University, January, 1980.

[20] Computational Understanding: Analysis of sentences andcontext, by CK Riesbeck. Working Paper, Fondazione Dalle Molleper gli studi linguistici e di communicazione internazionale,Castagnola, Switzerland, 1974.

[213 Scripts. Plans. Goals, and Undpratandinge An inquiry into* human knowledge structuren,by R Schank and R Abelson. Lawrence

Erlbaum Associates, Hillsdale, New Jersey, 1977.

(221 Understanding Goal-Based Stories,by R Wilensky. ResearchReport 140, Department of Computer Science, Yale University,September, 1978.

13

attempts to use the system in a new domain. A preliminary studyusing the general techniques advocated by the Yale group forparsing free text fields in formatted messages was completed in1979 and indicated evident applicability.

6 Planning and Problem Solving

Planning and problem solving -- the process of determining,examining, and deciding among alternatives -- is at the heart ofthe command and control domain. Knowledge representation andpresentation, natural language interfaces, and inference systemsare all useful as components to support the assessment anddecision processes.

Current artificial intelligence planning systems combineaspects of the preceding areas to propose action sequences foraccomplishing goals. Most of the existing systems also have someability to monitor the execution of proposed action sequences andto modify the plan adaptively to insure satisfactory achievementof the goal.

C2 planning systems could be used to support decisionmakingby providing an independent source of possible courses of action,accompanying justifications, and assistance in monitoring theexecution of selected courses of action. Planning in the C2environment obviously requires extensive use of domain-specificknowledge such as that cast in the chosen knowledgerepresentation. Ability to work with a powerful representation isa desirable feature. More important is the ability to cope withplanning conflicts such as unsatisfied preconditions, limitedresources, and counterplanning by other forces. Where possible, aC2 planning system should be able to "learn" common sequences ofactions for coping with problems of these types so that costlyreplanning need not be performed for commonly encounteredproblems. Execution monitoring is also a very desirable feature.

6.1 STRIPS, ABSTRIPS, and NOAH

STRIPS [233, ABSTRIPS [24], and NOAH (251 comprise a related

(233 STRIPS: A new approach to the application of theoremproving to problem solving, by RE Fikes and NJ Nilsson.Artificial Intelligence 2:189-208, 1971.

[24] Planning in a Hierarchy of Abstraction Spaces, by EDSacerdoti. Artificial Intelligence 5:115-135, 1974.

[253 A Structure for Plans and Behavior, by ED Sacerdoti.Elsevier, 1977.

14

IL • •

group of planning systems developed at SRI (and Stanford),originally for planning robot movement sequences. STRIPS, theoriginal system, used a predicate calculus representation of theworld and the actions available to the robot. Finding a sequenceof actions that achieved a goal was recast in this formalism tofinding a proof (in the sense of. logic) of the goal statedescription from the initial state description, by using theaction descriptions as theorems. Theorem proving techniques,notably resolution, were utilized. Unfortunately, this approachwas subject to strong effects from combinatorial explosion, whenthe size of the theorem-proving space grew much faster than thelength of the desired solution path.

A later modification to STRIPS permitted a form of learningin which action sequences that were commonly used could be savedand later used in planning as though they were single operations.The analogy with derived theorems as opposed to axioms in a logicsystem is very strong.

ABSTRIPS introduced another improvement to the basic STRIPSsystem, based on the idea of problem abstraction. Operators inSTRIPS had preconditions, a set of predicates which were requiredto be true before the operator could be applied. By ignoringpreconditions which were less "critical," ABSTRIPS could performa quick check to rule out many misleading action sequences beforeconsidering fine levels of detail at greater computational cost.

NOAH, while in the broad tradition of STRIPS and ABSTRIPS,is not a direct descendant of the other systems. The interestingidea here involves a procedural net, which links actions oroperators flexibly. In many earlier systems, selecting operatorscommitted the system to their order of execution. By representingthe actions as nodes in a network whose links are sequencinginformation, NOAH makes reordering of actions possible withoutcomplete replanning.

The availability of these systems is not known. A STRIPS-like system would be relatively easy to implement, ABSTRIPS moredifficult, and NOAH relatively complex. Good descriptions of allsystems exist in the literature.

The SRI systems have virtually defined the area of planningfor artificial intelligence systems. Any planning system must beable to deal with the problems that have been successfully facedby these systems and should attempt to tackle the problems thatare still open.

With the possible exception of NOAH, all the SRI systemsrely heavily on predicate calculus and theorem-provingtechniques. However, the underlying concepts are generally wellunderstood and documented, so adapting the approaches to otherrepresentations and inference engines is not impossible.

15

These planning systems have been used in a variety ofdomains, though never (so far as is known) in naval areas.Reliance on predicate calculus gives a uniform representation, sothat if naval tasks could be expressed in that way, presumablythese systems could then do planning in a naval domain. Apreliminary study effort is underway to examine the applicabilityof these approaches.

6.2 TALESPIN

A somewhat different approach to planning is embodied in theTALESPIN (26] program, out of Yale and UC Irvine. While manyunderlying concepts -- such as preconditions, actions, andpostconditions -- are shared with the SRI work, the motivation isquite different, and some critical implementation issues are alsoapproached differently.

TALESPIN is an effort to model (crudely) the planningprocesses that people use in problem solving. One constraintimposed by the model of people used by TALESPIN is that differentrepresentations be used for different kinds of knowledge; thusdifferent inference mechanisms are required. This is in contrastto the SRI approach which uses a uniform representation(predicate calculus) and inference mechanism (theorem proving).TALESPIN, like other Yale work, is in UCILISP on a non-ARPAnethost. Transfer and translation are required.

The vastly different outlooks represented by TALESPIN andthe SRI systems make meaningful comparison of the systemsextremely difficult. From some viewpoints they appear identical,while from others they are radically different. TALESPIN has notbeen used for any sort of naval domain problem solving to date.However, its insistence on representations and inferencemechanisms tailored to the information involved could prove to bea great aid in developing a naval-orieated system. TALESPINshares much philosophy and representation with the Yale naturallanguage efforts, which are promising. On the other hand, manyof the planning problems addressed by the SRI systems have notyet been considered in TALESPIN, most commonly because they havenot arisen in the course of TALESPIN's planning. A current studyis examining the TALESPIN approach and the naval domain forapplicability.

(261 The Metanovel: Telling Stories by Computer, by JR Meehan.PhD thesis, Yale University, December, 1976.

16

6.3 Other SRI Systems

Since the development of the various STRIPS systems andNOAH, work has continued at SRI in the area of planning. Currentefforts include (1) duveloping a representation for plans andactions that will ease user interaction for interactive plandevelopment [27], and (2) an attempt to perform planning in acommunity of actors [281, which involves modelling the knowledgeand actions of others as well as yourself. Most of this work hasjust started and is still somewhat in a design phase. However,such systems as have been constructed are presumably on theARPAnet.

This work addresses two issues that will play an importantrole in any command and control application of AI planningtechniques. Support of user interaction and direction iscritical, as is consideration of the plans and actions of others.However, it would seem that neither of these efforts is yet at auseful product stage. Further, they both rely on current SRIrepresentation techniques, which may differ from those used inother naval applications. SRI personnel have expressed interestin using naval domains for planning.

[27] Representing Knowledge in an Interactive Planner, by AERobinson and DE Wilkins. In Proceedings of the First AnnualNational Conference on Artificial Intelligence, pages 148-150,1980.

[281 Multiple-Agent Planning Systems, by K Konolige and NNilsson. In Proceedings of the First Annual National Conferenceon Artificial Intelligence, pages 138-141, 1980.

17

|A

7 Original List and Preliminary Evaluations

Languages

micro-PLANNER MIT BBACONNIVER MIT BBASCHEME MIT BBBPROLOG Waterloo BBA-FOLISP Stanford BCBPLASMA MIT CB-BETHER MIT CB-C

Knowledge representation systems (including someother "languages")

UNITS Stanford AB+A-KL-ONE BBN AB+A-KRL XEROX-PARC CBBFRL MIT BBA-KNOBS MITRE AA-BSNePS Buffalo B-BB+Hendrix SRI ?BB

Knowledge management systems - somewhat differentthan knowledge representation, though there iscertainly considerable overlap. Roughly analogousto DBMS vs. file design.

KBMS Stanford ?B-B-KLAUS SRI CBB-CYRUS Yale BB-B

Knowledge presentation systems - once you knowsomething, how do you share the knowledge?

AIPS BBN A-ABSDMS CCA ABA-

Inference systems (primarily - some also address otherissues)

E-MYCIN Stanford A-B-A-Hearsay ISI??) ?B+A-PROSPECTOR SRI ?BAROSIE RAND AA-A-RITA RAND ACACSA Maryland BBB

*See next page.

18

Natural Language Processing

ATNs Everywhere ABALIFER SRI B-BAYale work (ELI, etc.) BA-A-PARSIFAL MIT ?BB+

Planning and problem-solving

STRIPS SRI B-BAABSTRIPS SRI B-BANOAH SRI B-BABUILD MIT ?BBHACKER MIT ?BBGPS CMU ?B-ATALESPIN Yale/UCI B+A-BOther systems SRI BB+B-

Distributed AI and problem solving

Contract nets MIT, DREA BBBFA/C U Mass B-BBDSN Everywhere B-BB-

Database interfaces

LADDER SRI ABATED SRI AB+BTEAM SRI CB+C

Miscellaneous

Speech work various C+A-BBKG CMU ?CA-CHESS 5.0 NWU CCA-Vision work various ?CBTheorem proving various BBB+

*An explanation of grading:

Availability ---XXX--- Maturity

Applicability

Availability: A - On the net, in a usable form.Trivial acquisition.

B - Some minor problems; for example,not on net, or wrong LISP dialect.

C - Enough problems to be aninfluence on decisions.

19

Applicability: A - Obviously applicable.B - Probably useful, but not sure.C - Inapplicable.

Maturity: A - Mature. Techniques are well understood bya majority of practitioners. Limitationsas well as benefits are known.

B - "Adolescence"--generally known but withimportant gaps in knowledge, especiallyin demarcation of powers and limits.

C - "Infancy"--approach is new; little ifany established theory or past practice.

These scales are, it is hoped, orthogonal, and no judgmentshould be based on any one (except for elimination ofinapplicable techniques). For example, a nearly mature techniquewhich is hard to acquire may be less useful than a lessdeveloped, more accessible one. Also, a high score doesn't imply"goodness." A technology that can be shown to do only one thingcan be very mature.

20

8 Other Systems Suggested for Inclusion

Inference/Representation

KAS/Prospector SRIRLL StanfordAGE StanfordOPS5 CMUHEARSAY iSIEXPERT RutgersHASP SCI

Knowledge presentation

GUIDON Stanford

Database interfaces

FQL UPenn

9 Conclusions

The sixteen artificial intelligence systems summarized herewere felt to have some possible relevance to command and controlafter a preliminary survey. Deeper analysis has borne out thispresumption, though it has also shown that many of the systemsare still in early research stages and cannot be consideredfully mature. Most of the systems have not been applied tospecific problems in the Navy domain, so the evaluation ofrelevance is still preliminary.

10 Recommendations

The summaries of systems in this report should be used asintroductions for workers seeking to apply unfamiliar artificialintelligence technology to Navy problems. The evaluations shouldbe used to direct research into the application of thistechnology to command and control and to focus transitionefforts. Additional system evaluations should examine the othersystems suggested before seeking out new systems. i

DTII