Polyscheme: A Cognitive Architecture for Integrating Multiple Representation and Inference Schemes

Cassimatis, N. "Polyscheme: A Cognitive Architecture for Integrating Multiple Representation and Inference Schemes"


In order to understand and create human-level intelligence I have developed the Polyscheme cognitive architecture to build systems that combine several representation and inference schemes when they think.

Polyscheme is based on three principles. First, different (aspects of) situations that intelligent systems must deal with are best modeled with different schemes for representing knowledge and making inferences. Polyscheme includes several "specialists" such that each models a particular aspect of the world with its own (possibly unique) representation and inference techniques. Second, specialists must communicate with other specialists frequently so that each specialist uses the most complete, accurate and relevant information when it deals with a situation. Specialists in Polyscheme communicate and combine information by simultaneously concentrating on the same focus of attention. Finally, because information about some aspects of a situation is more important than information about others and because the order that specialists focus on those aspects is important, a system of focused specialists must have mechanisms that decide where to focus. Polyscheme's specialists, especially the reflective specialist, guide the focus of attention and thereby implement inference schemes using an "attraction" mechanism to specify their preferred foci.

Polyscheme enables multiple inference techniques to be integrated in dealing with a situation because each inference technique can be implemented with one or more focus schemes. I describe how to implement several important inference techniques (e.g., script matching, backtracking search, reason maintenance, stochastic simulation and counterfactual reasoning) as focus schemes.

I have used Polyscheme to implement the S6 system for common sense physical reasoning. S6 views interactions in a simple physical world through a 2-dimensional projection of that world. S6 keeps track of the identity of objects, infers the character and existence of events it cannot see, predicts the outcome of events, explains events and nonevents and revises its inferences when it receives new information. S6 successfully reasons about many scenarios researchers present to infants and young children in order to study their knowledge of the physical world.

S6 combines specialized representation and inference techniques for identity, time, events, causality, space and paths to successfully deal with a wide range of situations. The knowledge representation schemes S6 uses include scripts, frames, logical propositions, neural networks and constraint graphs. The inference schemes S6 implements include script matching, rule matching, backtracking search, neural network propagation and counterfactual reasoning. I show that these representation and inference schemes form part of a common sense substrate that underlies much of human cognition. The success of S6 therefore demonstrates that Polyscheme is important for understanding and building intelligent systems in any domain of human cognition.

Related Content