7/28/2019 OpenCogPrime

1/9

OpenCogPrime: A Cognitive Synergy Based

Architecture for Artificial General IntelligenceBen Goertzel, Novamente LLC, 1405 Bemerd Place, Rockville MD 20851, Email: ben@goertzel.org

Abstract- OpenCogPrime (OCP), a comprehensivearchitecture for artificial general intelligence (AGI) is brieflyoverviewed. Aimed in the long term at AGI at the humanlevel and beyond, the current partial implementation of OCPis being used for applications such as controlling virtual petsin virtual worlds and inferring novel conclusions from sets ofsemantic relations extracted from natural language. The keyaspects of OCP are described here in the context of thetheoretical foundation of "cognitive synergy theory"; and thecurrent implementation status is briefly reviewed.

I. INTRODUCTION

OpenCog Prime (OCP) is an architecture foradvanced artificial general intelligence, based on CognitiveSynergy Theory (CST) [1] and related ideas regarding thesystemic organization and dynamics of mind [2]. It isdescribed in a 300-page online wikibook [3] that iscurrently being transformed into an ordinary book for

publication in traditional paper form; our goal here ismerely to briefly summarize the architecture and its coreunderlying ideas.

OCP is closely related to the Novamente CognitionEngine (NCE) which has been reviewed previously [2,4,5],and has inherited some code as well as ideas from the latterarchitecture. The purpose of this paper is not to discuss thedifferences between OCP and NCE (which are mainly lowlevel and technical), but rather to articulate the conceptualfoundation of this family of AGI designs in the systemstheory of mind (and particularly in the notion of cognitivesynergy), more clearly than has been done before.

Our current plans for developing and teaching OCP

based systems largely center around embodying them inonline virtual worlds; however, that aspect of the projecthas been discussed extensively elsewhere [6,7] and will notbe the focus here, except for some illustrative examples.

One way to position this work in the general contextof cognitive informatics is in terms of the LRMB model[8] which explains the functional mechanisms andcognitive processes of natural intelligence in terms of 37cognitive processes at six layers known as the sensation,memory, perception, action, metacognitive, and highercognitive layers; and his concept of an autonomiccomputing system [9]. The breakdown of cognitiveprocesses we use in OCP is not identical to Wang's 37

process breakdown, but our breakdown could be mappedinto his with modest effort. In Wang's language, what we

Prlc.lib IEEE lI t ell '. II elll idlell 'lr.adcs (leel'09)I.Bacil, I.Wall, 1.1.lal, W.lilsler, I . ehal &L.A. Zadeh (Eds.)911-1-4244-4642-1/09/825.0020091EEE

60

the "cognitive synergy" principle underlying OCP posits isthat in order for an autonomic system to display a highlevel of general intelligence in the environments ofeveryday interest to humans, a great deal of dynamic,adaptive synergy between the multiple cognitive processesoperating in the system is required.

II. COGNITIVE SYNERGY THEORY

Cognitive Synergy Theory adopts a working definition ofintelligence as "the ability to achieve goals inenvironments", where the issue of how to weight differentgoals and environments is admitted as a subtle one. [10]argues that if one weights more highly the goals andenvironments relevant to existence in a community ofembodied communicative agents, then one obtainsnontrivial constraints on the sorts of systems that are likelyto be highly generally intelligent given feasiblecomputat ional resource restrictions. Specifically, it is

argued that this weighting leads naturally to cognitivearchitectures that contain multiple distinct but interactingtypes of memory, corresponding to the following types ofknowledge: declarative, procedural, sensory, episodic,attentional and intentional.The first four of these memory types are standard in

cognitive science [11,12]. Attentional knowledge isknowledge pertaining to which entities within the systemshould get space and time resources at a given moment;this is closely related to the notion of consciousness [13].Intentional knowledge refers to the system's overall goalsand derived subgoals (which may be continually revised bythe system's activity).

The essence of Cognitive Synergy Theory (CST) is thehypothesis that, in order to achieve a high level of generalintelligence in the context of a community of embodied,communicative agents, a system needs to:

Contain cognitive processes (Le. knowledgecreation mechanisms) specialized for each of theabove knowledge types

Contain methods for synergy between theseprocesses, so that the processes specialized foreach knowledge type can appeal to processesspecialized for other knowledge types for aid asneeded, often achieving dramatic efficiencyincreases as a result

CST divides cognitive processes into the two categories ofanalysis versus synthesis, and posits a "cognitive

7/28/2019 OpenCogPrime

2/9

schematic" that models the overall goal-achieving activityof an intelligent system; these aspects will be reviewed alittle later.

III. OPENCOG PRIME

Memory OpenCogPrime data st ructureTypeDeclarative The AtomTable, which is a special form

of weighted, labeled hypergraph -- i.e, atable of nodes and links (collectivelyreferred to as Atoms) with different

types, and each weighted with a multi-dimensional truth value (embodying an"indefinite probability" value that giveboth probability and confidenceinformation [15]). See [14] for a reviewof the system of Atom types.

Attentional Atoms in the AtomTable are weightedwith AttentionValue objects, whichcontain both ShortTermImportancevalues (governing processor timeallocation) and LongTerm Importancevalues (governing memory usage),

Procedural This is handled using special "Combo"tree structures embodying LISP-likeprograms, in a special program dialectintended to manage behaviors in avirtual world and actions in theAtomTable

Sensory Handled via a collection of specializedsense-modality-specific data structures

Episodic Handled via an internal simulation worldthat allows the system to run "mind'seye movies" of situations it remembers,has heard about, or hypotheticallyenvisions.

Intentional Goals are represented by Atoms storedin the AtomTable; there is a separatetable indicating which Atoms are top-level goals , which is used to guideattention allocation and goal refinementprocesses

Table 1. The OpenCogPrime data structures used to represent the keyknowledge types involved in Cognitive Synergy Theory

Tables 1 and 2 present the key s tructures and processesinvolved in OCP; and are ideally studied together withFigure 1 from [1]. Tables 3 and 4 exemplify thesestructures and processes in the application context of

embodied virtual agent control. Table 1 shows the

structures used in OCP for handling the key memory typesdiscussed in CST:

Table 2 shows the key cognitive processes considered inCognitive Synergy Theory, and then identifies the specificOpenCogPrime algorithms that embody these processes.Each of these cognitive processes deals with one or moretypes of memory -- declarative, procedural, sensory,episodic or attentional. . .

Next, CST defines an implication called the "cognitiveschematic" which serves as a general formulation of anintelligent system's basic cognitive activity:

Context & Procedure ~ Goal

This formula may be interpreted to mean "If the context Cappears to hold currently, then if I enact the procedure P, Ican expect to achieve the goal G with certainty p." Thesystem is initially suppl ied with a set of high-level goalssuch as "get rewarded by my teacher", "learn new things"and so forth; and it then uses inference (guided by othercognitive bemechanisms) to refine these initial goals intomore specialized subgoals.

In the OCP context, a procedure in this schematic is aCombo tree stored in the system's procedural knowledgebase and a context is a (fuzzy, probabilistic) logicalpredicate stored in the AtomTabIe, that hoIds, to a certainextent, during each interval of time. A goal is a fuzzylogical predicate that has a certain value at each interval oftime, as well.

The cognitive schematic leads to a partitioning ofcognitive processes into analysis versus synthesisprocesses. Synthesis processes ai"! to create. newimplications of the given form. Analys is processes ann toassess the truth values of implications of the given form.

CST suggests that social, communicative embodiedintelligence requires both analysis and synthesis processescorresponding to all the key knowledge types; and thateach of these processes must appropriately synergize withprocesses corresponding to other knowledge types.

Table 3 sums up the role of the different OCP cognitivemechanisms in terms of the CST categories of analysis andsynthesis.

As mentioned earlier, attentional knowledge is handledin OCP by the ECAN artificial economics mechanism, thatcontinually updates ShortTermImportance and LongTermImportance values associated with each item in thesystem's memory,

Cognitive OpenCogPrime algorithmProcessUncertain Probabilistic Logic Networksinference (PLN), a logical inference

framework capable of uncertain

reasoning about abstract knowledge,everyday commonsense knowledge,

61

7/28/2019 OpenCogPrime

3/9

and low-level perceptual and motorknowledge [15]

Supervised MOSES, a probabilisticprocedure evolutionary learning algorithm,learning which learns procedures (represented

as LISP-like program trees) based onspecifications [16]

Attention Economic Attention Networksallocation (ECAN), a framework for allocating

(memory and processor) attentionamong items of knowledge andcognitive processes, utilizing asynthesis of ideas from neural

networks and artificial economics.ECAN also comes with a forgettingagent that either saves to disk ordeletes knowledge that is estimatednot sufficiently valuable to keep inmemory. [15]

Map formation Use of frequent subgraph mining,MOSES and other algorithms toscan the knowledge base of thesystem for patterns and thenembodying these patterns explicitlyas new knowledge items

Concept A collection of heuristics for forming

creation new concepts via combining existingones, including conceptual blending,mutation and extensional andintensional logical operators

Simulation The running of simulations of(remembered or imagined) external-world scenarios in an internal world-simulation engine

Goal refinement Transformation of given goals intosets of subgoals, using conceptcreation, inference and procedurelearning

Table 2. Key cognitive processes in Cognitive Synergy Theory, and thealgorithms that play their roles in OpenCogPrime.

Synthesis Analysis

Declarative Evolutionary,blending andlogical conceptformation

PLN PLN forward PLN backward(Decl. an d inference inferenceProc.)

MOSES MOSES and Probabilistic(DecI. and hillclimbing modeling toProc.) procedure identify

learning patterns(combining amongportions and programsaspects of prior fulfilling aprocedures) certaion goal

in a certaincontext (partof MOSES)

Sensory/ Imagination of Filling in gapsEpisodic hypothetical in remembered

episodes based on orspecified criteria, hypothesizedvia combination episodesof aspects ofknown episodes

Attentional Hebbian learning Assignment ofImportance creditspreadingMap formation

Intentional Goal synthesis Goalrefinement

Table 3. The key OpenCogPr ime cognit ive pr ocesses catego ri zedaccording to knowledge type and process type

which control the amount of attention other cognitivemechanisms pay to the item, and how much motive thesystem has to keep the item in memory. HebbianLinks arethen created between knowledge items that often possessShortTermImportance at the same time; this is OCP 'sversion of traditional Hebbian learning.ECAN has deep interactions with other cognitive

mechanisms as well, which are essential to its efficientoperation; for instance, PLN inference may be used to helpECAN extrapolate conclusions about what is worth payingattention to, and MOSES may be used to recognize subtleattentional patterns. ECAN also handles "assignment ofcredit", the figuring-out of the causes of an instance ofsuccessful goal-achievement, drawing on PLN andMOSES as needed when the causal inference involved herebecomes difficult.The synergies between OCP's cognitive processes are

well summarized in Table 6 below, which is a 16x16matrix summarizing a host of interprocess interactionsgeneric to CST.

One key aspect of how OCP implements cognitivesynergy is PLN's sophisticated management of theconfidence of judgments. This ties in with the wayOpenCog's PLN inference framework represents truthvalues in terms of multiple components (as opposed to thesingle probability values used in many probabilisticinference systems and formalisms): each item in

62

7/28/2019 OpenCogPrime

4/9

OpenCog's declarative memory has a confidence valueassociated with it, which tells how much weight the systemplaces on its knowledge about that memory item. Thisassists with cognitive synergy as follows: A learning

mechanism may consider itself "stuck", generallyspeaking, when it has no high-confidence estimates aboutthe next step it should take.

Knowledge Type Virtual A2ent Example(s)Declarative The red ball on the table is larger than the blue ball on the floor

Bob becomes angry quickly Ball roll. Blocks don't. Jim knows Bob is not my friend.

Procedural A procedure for retrieving an item from a distant location A procedure for spinning around in a circle A procedure for stacking a block on top of another one

A procedure for repeatedly asking a question in different waysuntil an acceptabIe answer is obtainedSensory The appearance of Bob's face

The specific array of objects on the floor under the tableEpisodic The series of actions Bill did when he built a tower on the floor

yesterday The episode in which Bill and Bob repeatedly threw a ball back

and forth between each other The series of actions I just took, between getting up from the

chair and Bob saying "good"Attentional The set of objects that seem to be important in the context of

the game Bob and Bill are playing The set of words and phrases that are associated with Bob

being happy with me while we walk around togetherIntentional The goal of making Bob say positive things

The goal of making a tower that does not fall down easily The goal of getting Jim to answer my question

Table 4. Examples of the key knowledge types in the context of virtual agent control

63

7/28/2019 OpenCogPrime

5/9

CognitiveProcess ViriualA2entExampleInference Tall thin blocks, when stood upright, are less likely to topple

over if placed next to each other Bob hates cursing, and Jim is Bob's friend, and friends often

have similar likes and dislikes, so Jim probably hates cursingProcedure Learning Learning a procedure for crawling on the floor, based on

imitation of what others do when they describe themselves as"craw ling", plus reinforcement from others when they findone's imitation accurate

Learning a procedure embodying some combination offunctional and visual features that predicts whether someentity is considered a toy or not

Attention allocation Pictures of women are associated with Bob's happiness, andBob's happiness is associated with getting reward, therefore

pictures of women are associated with getting reward Asking for help is surprisingly often a precursor to gettingreward when Jane is around; so when a reward is gotten whenJane is around, a little extra attention should be given toongoing improvement of the processes that help in themechanics of asking for help

Goal refinement The goal of making Jim happy, seems to often be achieved bythe goal of creating sculptures Jim likes, and Jim likescomplicated sculptures; thus I adopt the goal of creatingcomplicated sculptures when Jim is around

Declarative pattern mining Tall thin blocks, when stood upright, are likely to topple overSensory pattern recognition When Jim builds a castle out of blocks, he identifies some

portions of the castle as "towers" and others as "walls"; it's

necessary to visually identify which portions of each castlecorrespond to these descriptors

It's also necessary to visually identify the castle as a wholeversus the tabIe,floor or other base it's resting on

Simulation Using an internal simulation world to experiment withbuilding various towers rapidly, at a pace faster than ispossible in the online simulation world where humansparticipate

Using an internal simulation world containing a simulation ofBob and Jim, to simulate what Bob will know about whatyou're doing if you hide behind Jim and build a tower ofblocks

Concept creation The concept of an unstable structure The concept of an irritable person The concept of a happy occasion

Map formation The set of all knowledge items associated with Bob being in agood mood (which may then be used to form a new concept)

The set of all knowledge items associated with (running,walking or crawling) races

Table 5. Examples of the key cognitive processes in the context of virtual agent control

64

7/28/2019 OpenCogPrime

6/9

How--> Ma p fCM'1T1ation Goal sy s tem Simulation SensorimotcwHelps

1pattern nH:ognition

\11

Uncer tain infer-.nee Creates new concepts Goal refinement - Simulations provide Creates new conceptsand relaoonships. enables more careful a method of testing an d relationships.enabl ing br ieferusefu l goal-based inference speaJlative enabling b riefer usefulinference traits pruning infet'ential inference trails

COndUSKlnS- Simulationssuggest hypothesesto be explored viainference

Supervised Creates new Goatref inement al lows Simutation provides Extraction of sensorimotorp rocedu r e lea rn ing proceduras to be used more precise defin i tion a method of "fitness pat terns al lows creat ion of

as modules in of fi'tness functions, estimation .. allowing abstracted fitnesscandidate procedures making procedure inexpensive testing functions for (inferentially

leaming's jo b eesioer of candidate and simulative'y)procedures eYaluating procedures

gu id ing real -work l act ions

Attention allocation Creates ne w concepts Goal r ef in emen t a l l ow s S imul at il Xl p ro vi de s Creates conceptsgrouping ..attentionally mo re accu r at e ly go al - datafor attention grouping ..a ttentionaltyre'ated" memory items. driven allCJcation of alklcation - allowing related" memory items.enabling AA to find attention attentionsl enabling AA to find subtlersubtler attentional information to be attentionsl patternspatlEH'ns ilwotving these extracted from co - involYing these nodesnodes occLWences

observed il lsimulation

Concep t c r eat ion Creates new concepts Goal refinement Utility of concepts Creates new concepts toto be fed into o1het' prov ides more precise ma y be assessed via be fed into other conceptconcept creation defillition of criteria via aeating s im ula ted c reati on m ec ha ni sm smechanisms which new concepts entiti>es embodying

are ae.ated th e new conceptsand seeing whatthey lead to insimulation

How--> Uncerta.n Inference Supervised procedure- Attention aJaocatkHI Concept cre-anonHelps 1 leamlng

\11

UncertajnInference NA When Inference gets Importance le... els allow ProvKjes new concepts,stuD: In an Inferenc.e pruning o-f inference alJowlng briefer usefultrail, tt can ask procedure trees Inference traitstNmlng 10learn newpatterns regardingconcepts In the Inferencetrail (if there ls adequatedata re-gardlnQ theconcepts)

Supervised procedure- Inference can be used to NA Importance le... els may ProvKjes new concepts,leamlng alJowprior expenence to be used to bias tholces alJowlng compacter

guide each Instance of made In 1 he course of programs uSing n ewprocedure learning. procedure learning concepts In vanous roles

{e.g. In OCP. in theiltness evaluation andrepresentatlon-DulldlDgphases o-fMOSES)

Attention aJaocatkHI Enables Inference o-fnew Procedure learning can NA Ca-mbfiaoon of conceptsHebbianLinks a nd recognize patterns In formed ...a mapHebbianPredlcates fra-m htstorical syslem activity. formation I may lead to-existing ones whICh are then used to new concepts that eyen

oolld concepts and better dire-c:t anenuonrelationships gUldngatlantlon allocatioo

Concept cre-anon AlJows inferenbal Procedure learning can Allows assessment of NAassessment o-fthe yslue be used to search fur the value o-fnewof new concepts tvgh-quallry blends of concepts based on

exist ing concepts (using hts .toricalat tenbonate.g. Inferential and 1mowledgeatlentlonal knowledge Inthe fltnes.sfuncuons)

65

7/28/2019 OpenCogPrime

7/9

How--> UncertaJnInference Supervlaed procedure Attention aJlaeatkm Concept creationHe.pl 1 learning

\11

Map formation Speculative Procedurelearningcan Att'9ntlon allOCatJon No SJgnlflc.sntOJrSctInference [an help be use

7/28/2019 OpenCogPrime

8/9

Without reasonably accurate confidence assessment to guideit, inter-component interaction could easily lead to increasedrather than decreased combinatorial explosion. And of coursethere is an added recursion here, in that confidence assessmentis carried out partly via PLN inference, which in itself reliesupon these same synergies for its effective operation.To illustrate this point further, consider one of the

synergetic aspects described in TabIe 3 in [1]: the rolecognitive synergy plays in deductive inference. Deductiveinference is a hard problem in general- but what is hard aboutit is not carrying out inference steps, but rather "inferencecontrol" (Le., choosing which inference steps to carry out).Specifically, what must happen for deduction to succeed inOCP is:

1. the system must recognize when its deductiveinference process is "stuck", i.e, when the PLNinference control mechanism carrying out deduction

has no clear idea regarding which inference step(s) totake next, even after considering all the domainknowledge at is disposal

2. in this case, the system must defer to another learningmechanism to gather more information about thedifferent choices available - and the other learningmechanism chosen must, a reasonable percentage ofthe time, actually provide useful information thathelps PLN to get "unstuck" and continue thedeductive process

For instance, deduction might defer to the "attentionalknowledge" subsystem, and make a judgment as to which ofthe many possible next deductive steps are most associated

with the goal of inference and the inference steps taken so far,according to the HebbianLinks constructed by the attentionallocation subsystem, based on observed associations. Or, ifthis fails, deduction might ask MOSES (running in supervisedcategorization mode) to learn predicates characterizing someof the terms involving the possibIe next inference steps. OnceMOSES provides these new predicates, deduction can thenattempt to incorporate these into its inference process,hopefully (though not necessarily) arriving at a higherconfidence next step..

IV. CURRENTLY IMPLEMENTED COMPONENTS

OpenCogPrime is currently partially implemented within abroader C++ software framework called OpenCog [16], whichwas created during 2007-2008 with a view toward broadlyfostering integrative open-source AI software development, aswell as providing a software foundation for OCP. OpenCog isnow supported by a small but active developer community.In addition to the OpenCog framework itself, some of the keyOCP components already substantially implemented are:

A version of the PLN probabilistic logical inferenceengine (which however has some lacunae, e.g it doesnot include the full PLN rules for spatiotemporal orintensional inference, and its adaptive inferencecontrol mechanisms are incomplete)

A version of the MOSES probabilistic evolutionaryprogram learning algorithm (which however

67

currently handles only a subset of the space ofpossible programmatic constructs: it deals with loopsand conditionals and Boolean and real inputs, but notyet internal variables or higher order functions)

A version of ECAN; and also, a simple forgettingmechanism that removes Atoms with low long-termimportance

Proxies between OpenCog and the RealXTend andMultiverse virtual worlds, used for controlling virtualdogs and humanoids

An initial goal system and procedure executionframework, used for executing learned procedures inthe above virtual worlds

V. CONCLUSION

To thoroughly describe a comprehensive, integrative AGIarchitecture in a brief conference paper would be animpossible task; all we have attempted here is a briefoverview, drawing heavily on the formulation of CognitiveSynergy Theory from [1] to structure the discussion. We donot expect this brief summary to be enough to convince theskeptical reader that the approach described here has a highodds of success at achieving its stated goals; but we will rateour summary successful if the reader is motivated to lookmore deeply by exploring [1,2,3,15] and other relatedreferences that provide more depth.

REFERENCES

[I] Goertzel, Ben. Cognitive Synergy: A Universal Principle of FeasibleGeneral Intelligence?, Dynamical Psychology, 2009

[2] Goertzel, Ben. The Hidden Pattern. BrownWalker, 2006[3] Goertzel, Ben. OpenCog Prime: Design for a Human-Level AGI.

Online at OpenCog.org/OpenCogPrime, 2009[4] Goertzel, Ben. Patterns, Hypergraphs and General Intelligence.

Proceedings of WCCI 2006, Vancouver.[5] Goertzel, Ben, Cassio Pennachin, Andre Senna, Thiago Maia and

Guilherme Lamacie. Novamente: An Integrative Architecture forArtificial General Intelligence. Proceedings of IJCAI-03 Workshop onAgents and Cognitive Modeling, Acapulco, August 2003

[6] Goertzel, Ben. A Pragmatic Path Toward Endowing Virtually-Embodied AIs with Human-Level Linguistic Capability, Special Sessionon Human-Level Intelligence, IEEE World Congress on ComputationalIntelligence (WCCI) Hong Kong, 2008

[7] Goertzel, Ben. What Must a World Be That a Humanlike AGI MightDevelop In It?, Dynamical Psychology, 2009

[8] Wang, Y., Y. Wang, S. Patel, and D. Patel. A Layered Reference Modelof the Brain, IEEE Transactions on Systems, Man, and Cybernetics (C),Vo1.36, No.2, March 2006, pp.124-133, 2006

[9] Wang, Y. Toward Theoretical Foundations of Autonomic Computing,The International Journal of Cognitive Informatics and NaturalIntelligence (IJCINI), IPI Publishing, USA, 1(3), July 2007, pp.1-16,2006

[10] Goertzel, Ben. The Embodied Communication Prior. ICCI-2009[II] Tulving and Craik. The Oxford Handbook of Memory. Oxford

University Press, 2005[12] Eichenbaum, Henry. The Cognitive Neuroscience of Memory. Oxford

University Press, 2002[13] Metzinger, Thomas (Editor). Neuronal Correlates of Consciousness.

MIT Press, 2005

[14] Goertzel, Ben and Cassio Pennachin. The Novamente AI Engine. InGoertzel and Pennachin, Artificial General Intelligence, Springer, 2002

[15] Goertzel, Ben, Matt Ikle, Izabela Freire Goertzel and Ari Heljakka.Probabilistic Logic Networks. Springer, 2008

7/28/2019 OpenCogPrime

9/9

[16] Looks, Moshe. Competent Program Evolution. PhD thesis, ComputerScience Dept. Washington University St. Louis, 2006

[17] Goertzel, Ben, Matt Ikle and Joel Pitt. Economic Attention Networks.Proceedings of AGI-09, Atlantis Press, 2009

[18] Hart, David and Ben Goertzel. OpenCog: A Software Framework forIntegrative Artificial General Intelligence, in Proceedings of the FirstAGI Conference, Ed. Wang et ai, lOS Press, 2008

68