Rethinking Cognitive Architecture! Reconciling Uniformity and Diversity via Graphical Models! Paul Rosenbloom!!! 1/25/2010! Department of Computer Science &! Institute for Creative Technologies! The projects or efforts depicted were or are sponsored by the U.S. Army Research, Development, and Engineering Command (RDECOM) Simulation Training and Technology Center (STTC). The content or information presented does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred.
Cognitive Architecture! Fixed structure underlying cognition! Defines core memories, reasoning processes, learning mechanisms, external interfaces, etc.! Yields intelligent behavior when combined with knowledge in memories! Including more advanced reasoning, learning, etc.! May model human cognition and/or act as core of a virtual human or intelligent agent! Strong overlap with intelligent agent architecture! 2!
Example Virtual Humans! 1997 Improving in mind and body over the years! These all based on Soar architecture! 2000 2009 3!
Soar Architecture (Versions 3-8)! Symbolic working memory! Long-term memory of rules! Decide what to do next based on preferences generated by rules! Reflect when canʼt decide! Learn results of reflection! Interact with world! Pursued both as a unified theory of cognition and as an architecture for virtual humans and intelligent agents 4!
Systems Levels in Cognition! In the large, architecture is a theory about one or more systems levels in an intelligent entity! Usually part of Cognitive Band! At each level, a combination of structures and processes implements basic elements at the next higher level! (Newell, 1990) 5!
Hierarchical View of Soar (Versions 3-8)! Scale! Functionality! Mechanism! Details! 1 sec! Reflective! 100 ms! Deliberative! 10 ms! Reactive! Problem Space Search! Decision Cycle! Elaboration Cycle! Impasse/Subgoal! If Canʼt Decide! Preference-based Decisions upon Quiescence! Parallel Rule Match & Firing! Learning by Chunking 6!
Level Girth! Range of structures & processes at a level! Key issue: uniformity versus diversity! Uniformity: Minimal mechanisms combining in general ways! Appeals to simplicity and elegance! The physicistʼs approach! Challenge: Efficiency and full range of required functionality/coverage! Diversity: Large variety of specialized mechanisms! Appeals to functionality and optimization! The biologistʼs approach! Challenge: Integrability, extensibility and maintainability! Across a hierarchy, level girth may stay comparable or vary! Physicists and biologists likely assume uniform! Can a mixture of some sort yield benefits of both?! Network researchers assume hourglass! 7!
The Internet hourglass Applications! Web! FTP! Mail! News! Video! Audio! ping! napster! Everything on IP Transport protocols! TCP! SCTP! UDP! ICMP! IP! Ethernet! 802.11! Power lines! IP on ATM! everything Optical! Link technologies! Satellite! Bluetooth! 7/24/09 Paul S. Rosenbloom From Hari Balakrishnan
What About Cognition?! Top (applications) is clearly diverse! Key part of what architectures try to explain! Bottom is likely diverse as well! Physicalism: Grounded in diversity of biology! Strong AI: Also groundable in other technologies! Is the waist uniform or diverse?! Hourglass or rectangle! Traditionally question about the architecture! Applications Architecture Implementation 9!
Architectural Uniformity vs. Diversity! Soar (3-8) is a traditional uniform architecture! ACT-R is a traditional diverse architecture! Soar 3-8 ACT-R 10!
Examples! Soar (3-8) is a traditional uniform architecture! ACT-R is a traditional diverse architecture! Recently Soar 9 has become highly diverse! Soar 3-8 Soar 9 11!
Towards Reconciling Uniformity and Diversity! Accept diversity at the architectural level! Move search for uniformity down to implementation level! Biological Band in humans! Locus of neural modeling! Computational Band in AI! Normally just Lisp, C, Java, etc.! Impacts efficiency and robustness but usually not part of theory! Base on graphical models! (Newell, 1990) Applications Architecture Implementation Computational Band! 12!
Graphical Models! Efficient computation with multivariate functions by decomposition into products of subfunctions! For constraints, probabilities, speech, etc.! u x Come in a variety of related flavors! Bayesian networks: Directed, variable nodes! E.g, p(u,w,x,y,z) = p(u)p(w)p(x u,w)p(y x)p(z x)! w Markov networks: Und., variable nodes & clique potentials! Basis for Markov logic and Alchemy! Factor graphs: Und., variable & factor nodes! E.g., f(u,w,x,y,z) = f 1 (u,w,x)f 2 (x,y,z)f 3 (z)! w Compute marginals via variants of! u Sum-product (message passing)! Monte Carlo (sampling)! f 1 x y f 2 z y z f 3 13!
Properties of Graphical Models! Significant potential for resolving diversity dilemma! State of the art performance across symbols, probabilities and signals via uniform representation and reasoning algorithm! Representation: My focus has been on factor graphs! Reasoning: Sum-product passes messages about elements of variablesʼ domains! Sum-product algorithm yields! (Loopy) belief propagation in Bayesian networks (BNs)! Forward-backward algorithm in hidden Markov models (HMMs)! Kalman filters, Viterbi algorithm, FFT, turbo decoding! Arc-consistency and production match! Many neural network models map onto them! Support both mixed (symbolic & probabilistic) and hybrid (symbolic & signal) processing! 14!
Mixed and Hybrid Processing! Symbol processing is the standard in cognitive architectures! E.g., given that birds fly and that Tweety is a bird, conclude Tweety flies! But much knowledge is uncertain! Even knowledge about whether birds fly! Tweety may be a penguin, or have a broken wing, or! Reasoning under uncertainty has been a major recent focus in AI! P(flies bird) =.95! Bayesian networks enable efficient computation with large probabilistic models! Only recently been combined effectively with general symbol processing, to yield mixed processing, and has had little impact on cognitive architectures! Virtual humans also need to perceive their worlds! Speech, vision, etc. require extended signal processing! Also significant progress via grahical models! Hidden Markov models for speech and Markov random fields for vision! Architectures view this as preprocessing that creates symbols! u x Need hybrid processing to use reasoning in perception and perception in reasoning! y p(u,w,x,y,z) = p(u)p(w)p(x u,w)p(y x)p(z x) 15! w z
Scope of Sum-Product Algorithm! Message/Variable Range Boolean Numeric Message/Variable Domain Discrete Continuous Symbols Probability (Distribution) Signal & Probability (Density) Mixed models combine Boolean and numeric ranges! Hybrid models combine discrete and continuous domains! Hybrid mixed models combine all possibilities! Dynamic hybrid mixed models add a temporal dimension! 16!
Research Goals and Plans! Long-term goals! Evaluate ability of graphical models to uniformly support existing architectures! Develop novel, more functional architectures! Enhancing and/or hybridizing existing architectures! Starting from scratch leveraging strengths of graphical models! elegance, functionality, extensibility, integrability, maintainability! Near-term strategy! Develop a mixed hybrid version of the Soar architecture! One of the longest standing and most broadly applied architectures! Both as unified theory of human cognition and architecture for virtual humans! Exists in uniform (Soar 8) and diverse (Soar 9) forms! 17!
Status! Developed memory core of a mixed decision cycle! Combines symbols and probabilities in service of decision making! Has nascent hybrid aspect, with continuous variables! Potential for uniform implementation of much of Soar9ʼs diversity! Working memory becomes a multivariate continuous function! Represented as an N dimensional piecewise linear function! Domains are discretized for integers! Ranges are also Booleanized for symbols (w/ symbol tables)! y x [0,10>! [10,25>! [25,50>! [0,5>! 0!.2y! 0! [5,15>!.5x! 1!.1+.2x+.4y! 18!
Status (2)! Rules become general conditionals! Conditions and actions embody normal rule semantics! Condacts embody (bidirectional) constraint/probability semantics! Encoded in graph and processed via sum-product algorithm! CONDITIONAL ConceptPrior P(c)! Condition: Object(s,O1)! Condact: Concept(O1,c)! Walker! Table! Dog! Human!.1!.3!.5!.1! Object: WM Constant Pattern Join Distribution Concept: 19!
Status (3)! Implemented four distinct memories! Procedural! Rule: If-then triggering of retrieval based on direct cues! Declarative! Semantic: Predict concepts and unknown object features from known object features (cues)! Episodic: Retrieve best matching past episode based on cues! Constraint: Compute value combinations within constraints! Integrated memories together! Mix procedural and declarative within conditionals! Use results from one memory as cues to others! 20!
Summary and Future! GMs show potential for resolving diversity dilemma! A uniform approach/implementation for architectural and representational diversity! Combining simplicity with broad functionality (and efficiency?)! Mixed and hybrid representation and processing! Mixed: symbolic reasoning under uncertainty! Hybrid: Merging perceptuomotor behavior with cognition! Working towards a mixed, hybrid variant of Soar 9! Leverage mixed processing for semantic and episodic memories! Leverage hybrid processing for imagery (and perception)! Much continued room to grow past Soar 9! Motor control, emotion, theory of mind, etc.! Radically new architectures! 21!