Polyscheme John Laird February 21, 2008

Major Observations Polyscheme is a FRAMEWORK not an architecture – Explicitly does not commit to specific primitives for knowledge representation and processing – Provides a control structure to integrate together processing modules for a given system/agent Integration is at the level of individual common functions – Instead of Soar’s PSCM functions (elaborate, proposal, evaluate, select, apply) – Polyscheme seems to have more varied functions and oriented around what is reasoned about (time, causality, events) instead of action. – Emphasizes multiple algorithms as opposed to multiple sources of knowledge – Attempts to have interaction between modules for individual steps – Segregates knowledge based on implementation medium Similar to Soar, but not based on sources of knowledge from learning. Soar has a fixed set of modules and has tighter integration of those modules – But does not provide a means to easily integrate new modules – Soar does not provide parallel integration – rules – impasses – memories …

Claims Cognitive substrate hypothesis: “a relatively small set of properly integrated data structures and algorithms can underlie the whole range of cognition required for human- level intelligence.” – Computational problems that need to be solved: “… reasoning about temporal intervals, causal relations, identities between objects and events, ontologies, beliefs, and desires.” – “… reasoning about time, space, part-hood, categories, causation, uncertainty, belief, and desire.” – “The cognitive substrate hypothesis suggests that once a set of computational mechanisms, in other words, a cognitive substrate, that solves the problems of human-level AI for this set of problems is constructed, achieving the rest of human- level AI will be a relatively easy problem.”

Claims 2 1.“Identifying and implementing a cognitive substrate will accelerate progress toward human-level intelligence.” – “Creating intelligent systems for new domains is accelerated.” – “It took very little extra work to design a parser once the structural dualities between syntax and physical law were found. No new algorithms need to be designed, thus making the procedural profusion problem less severe.” 2.“As work with Polyscheme illustrates, these approaches enable great progress in solving the integration problems associated with implementing a cognitive substrate.” – “…the problem of achieving human-level AI is reduced and simplified.” – “Instead of needing enormous databases of knowledge and hundreds or thousands of algorithms to achieve human-level intelligence, researchers can focus on solving the problems of human-level AI for a relatively small (but still difficult) set of problems knowing that other domains can be addressed by mapping them onto a substrate.”

Claims 3 Systems built using Polyscheme demonstrate that algorithmic hybrids implemented using a focus of attention can (1) exhibit more characteristics of intelligence than individual computational methods alone and (2) deal with problems that have formerly been beyond the reach of synthetic computational intelligence.”

Theoretical Picture Common FunctionsMultiple Implementations Individual and Hybrid Algorithms ]Forward Inference Subgoaling Simulate Alternate Worlds Identity Matching] [store information, offer opinion, forward inference, request information (subgoal) identity, represent alternative worlds] Cognitive substrate = “… reasoning about time, space, part-hood, categories, causation, uncertainty, belief, and desire.” Neural Networks Forward Rule Changing (?) Ontologies Search Stochastic Simulation Focus of attention fuses results of specialists’ inferences together and a set of attention- controlling strategies shape flow of computation. specialists

Key Ideas in Implementation Specialists – Implementations of specific methods/algorithms – Appear to be no constraint on how implemented. – “Polyscheme modules can have arbitrarily large, varied and persistent memories …” Communication between specialists – Propositional language – “Specialists must be able to translate back and forth into this language in order to inform other specialists of their inferences and receive knowledge from other specialists.” Focus of Attention – A proposition that all specialists focus on. Behave is a sequence of “attention fixations” (involving all specialists) called “focus traces”. “When an algorithm performs a common function on a proposition, we say it attends to or fixes its attention on this proposition.” No fixed control structure – But seems you “implement” one through controlling attention fixation. Execution of specialists are interleaved (serial and parallel) to process focus of attention.

Polyscheme Control Loop Do forever: – A = FM.getNextFocus(). – For all pairs of specialists (S1, S2) S1.store (P, S2.opinionOn(A)) – For all S S.modifyFM(A)

Evaluation What about – Computational resources required – Bounded execution – guaranteed reactivity – Required knowledge to generate/control behavior – Types of knowledge that can be used to generate/control behavior – Reliability – Ease of modification (and debugging) – Expressiveness – Learnability