Polyscheme

Principles | Framework and Architecture | Differences from other approaches
Human-Level Intelligence Laboratory | Cognitive Substrate 

Polyscheme is a cognitive architecture designed to achieve human-level artificial intelligence and to explain the power of human intelligence.  We choose and evaluate research efforts primarily on how they help advance the ability of systems to do what people can and computers can't yet.  Polyscheme has enabled computer models that explain how background knowledge and context can be used to understand language that is ungrammatical, nonliteral and ambiguous; robotic systems that can reason and plan while incorporating incomplete and noisy information from a dynamic environment; and computational models of children's physical, social and linguistic reasoning.

If you would like to get a copy of Polyscheme, please contact Nick Cassimatis.

Principles

Polyscheme is based on the belief that the following are key obstacles to human-level computational intelligence:

Current computer languages and data structures do not have the ability of humans to deal with ambiguous, incomplete, nonliteral and ill-formed language and knowledge representations. Much of human reasoning and learning involves finding the best model(s) of the world.  The fact that these models involve unknown objects, times, identities, spatial locations and counterfactual states means that there are an extremely large number of worlds to choose from. Many aspects of human-level intelligence are currently dealt with by known computational methods, but full human-level intelligence requires these to be integrated.  These data structures and algorithms appear very different and difficult to integrate.

The following principles underlie Polyscheme.

• Attention and Simulation.  The mechanisms of perception, attention and imagery interact to execute powerful algorithms based on mental simulation.  This paper argues this point.

• Common function.  There are small number of "common functions" into which reasoning algorithms required for human-level intelligence can be decomposed.

• Multiple implementation.  Multiple data structures and algorithms, based on very different representational formalisms, can be used to implement these common functions.

• Cognitive self-regulation. The control of cognition is conducted in large part by mechanisms used to correct cognitive problems.

• Cognitive Substrate.  A basic set of reasoning mechanisms for a set of relations (involving, e.g., time, space, identity, causality, parthood, categories and counterfactual worlds) underlie cognition in many domains.  This implies that cognition in, for example, natural language use and social interaction is not based on any mechanisms specific to language or social relations.  This paper gives a somewhat dated explanation of this point.

Framework and Architecture

The Polyscheme Framework (unfortunately often referred to as an architecture in papers through 2009) is an approach to integrating multiple computational mechanisms, which may or may not have any relation to the mechanisms of human cognition.  The Polyscheme Cognitive Architecture uses the Polyscheme Framework to integrate mechanisms we believe correspond to those in human cognition.

Differences from other approaches

AI algorithms and human cognition.  It is commonly thought that reasoning methods in AI are not closely related to the mechanisms of human cognition.  Our work is based on the the belief that in fact AI algorithms and the mechanisms of human cognition have many deep similarities.

Cognitive architectures and cognitive modeling

• Polyscheme is not a production system.  Cognitive architectures are commonly identified with rule-based production systems.  However, cognitive architectures are theories of what is invariant across cognition in many domains.  Polyscheme is based in part on the belief that multiple computational mechanisms are part of intelligence and is not a production system.
• Higher-level formalisms and mechanisms.  Polyscheme is based on data structures that are both more complex and more specific than those in many other architectures.  For example, many architectures base their representational formalisms on some variation of attribute/value lists.  This is very general and flexible, but requires either glossing over or spending a great deal of time to reinvent machinery for concepts involving, e.g., time, causality, categories and identity.  Polyscheme provides such mechanisms as part of the architecture and thus in many cases obviates a great deal of effort.  The Cognitive Substrate theory suggests that such mechanisms are sufficient to model cognition in a wide variety of domains.
• Higher-order cognition.  Polyscheme has been designed in large part to support what is often called "higher-order cognition", e.g., reasoning, problem solving and language use.
• Emphasis on ability, breadth and parsimony rather fitting quantitative data.  It is common in the cognitive modeling community to suppose that carefully modeling quantitative data is key to models of human cognition.  When the goal is human-level intelligence, however, it is important to focus on ability, breadth and parsimony.  Polyscheme modeling emphasizes these properties. 

"Modern" artificial intelligence

• Not normative.  Many AI approaches aim for some normative standards such as logical or probabilistic correctness.  Humans often depart from these standards while nevertheless performing impressively in many situations.  Thus, these normative standards are not a primary goal of our work.
• More general applications.  While AI methods are often in theory aimed at general problems, they are often applied to very specific problems.  Our goal is to reach the level of generality of human cognition.  This involves the ability to learn skills and knowledge and to use these in adapting to unforeseen situations. 
• Not fully general.  Most AI methods characterized quite generally.  We are happy to only achieve competence in those areas that people do.