Polyscheme Overview
Polyscheme | Human-Level Intelligence Laboratory

Objective

Polyscheme is a cognitive architecture designed to model and achieve human-level intelligence by integrating multiple methods of representation, reasoning and problem solving. 

Principles Underlying Polyscheme

• Procedural Substrate.  Most high-order reasoning and problem solving algorithms (even those from very different subfields of AI, such as logic-theorem proving and probabilistic inference) can be implemented using the same set of basic computational operation.  These include forward inference, subgoaling, grounding, representing alternate worlds and identity matching.
• Multiple representations.  Each basic operation can be implemented using multiple representations, including those arising from perceptual processes.
• Representational Substrate.  Cognition about a basic set of relations (involving times, space, events, identity, causality and belief) can underlie cognition in many domains.

• Specialists.  Specialists are modules based on a particular representation.  Each specialist executes each of the basic operations of the procedural substrate. 
• Integrative focus of cognitive attention.  All the specialists focus on the same aspect of the world simultaneously
• Attention control implements algorithms.  Policies for guiding the focus of attention implement high-order AI algorithms.  For example, the policy, "when uncertain about A, focus on the world where A and focus on the world where not-A" implements backtracking search" and "when A is more likely than not-A focus on the world where A more often than you focus on the world where not-A" implements stochastic simulation.)

• Integration of "high-level" and "low-level" cognition, perception and action.  Since all high-level reasoning and planning algorithms are implemented by a focus of attention that integrates all lower-level representations and perceptual and motor processes, every single step of reasoning and planning is guided by multiple representations and is informed by new perceptual information.  This is especially helpful in robotics and modeling physical reasoning, where a changing environment and noisy sensors can quickly invalidate plans and inferences.
• Integration of multiple higher-level cognitive processes with each other.  Very different reasoning algorithms, from truth-maintenance, backtracking search, stochastic simulation and logic theorem proving can interact to reason and solve problems in the same situation since they are each implemented using the same focus of attention.

How Polyscheme differs from other architectures

• Higher-level basic services than most cognitive architectures.  These include reasoning about events, time, space, causality, identity, desire and beliefs. Although this is more constraining than the more basic data structures in most architectures, it enables modeling of more complex cognition and, if the cognitive substrate principle is correct, is sufficient for reasoning in most domains
• No homunculus.  Most cognitive models completely determine which action should be taken in each situation.  Polyscheme modeling aims to understand how these actions are chosen in the first place and thus choose actions as the result of reasoning and problem-solving instead of a priori modeler decisions.
• Only one model.  Most cognitive architecture communities design a different model for each task or phenomenon.  This means that they might be inconsistent with each other and that modeling idealizations could abstract away from difficult problems.  In Polyscheme, the goal is to have accounts of reasoning multiple tasks as part of the same model so that integration is an unavoidable and constant process and so that it is more difficult to gloss over hard problems. Polyscheme does not commit modelers to a single representational formalism or input format.
• Multiple representations.  Polyscheme does not commit users to a single representational formalism.
• Language is an important focus of Polyscheme modeling, which has historically (and with a few exceptions) not been the case in most other modeling communities.

• A model of (mostly infant) physical reasoning, developed mostly at the MIT Media Laboratory.  I am currently writing an article on this.
• A robotics framework, developed at the Naval Research Laboratory, for using high-level, “good old-fashioned AI” algorithms with reactive robotic subsystems.  It is commonly thought that these algorithms make robots more brittle, but when integrated with reactive subcomponents using Polycheme, they not only make robots more flexible, but also able to achieve more complex goals.
  • N. L. Cassimatis, J. Trafton, M. Bugajska, A. Schultz (2004). Integrating Cognition, Perception and Action through Mental Simulation in Robots. Journal of Robotics and Autonomous Systems. Volume 49, Issues 1-2, 30 November 2004, Pages 13-23.
  • See the (44MB, working on a smaller version) video of this system working.  The goal of the robot is to track the orange robot it sees to the right of the scene in the beginning.  Notice how it revises its plan, in the middle of executing it, because of new sensor information and physical reasoning. 
• A model of syntactic parsing model based almost entirely on the mechanisms in the physical reasoning model, making the case for the cognitive substrate principle.
  • N. L. Cassimatis (2004). Grammatical Processing Using the Mechanisms of Physical Inferences.  In Proceedings of the Twentieth-Sixth Annual Conference of the Cognitive Science Society.  (pdf)
• A framework for answering queries (soundly and completely) over multiple sources of information and computation, each based on different computational methods. 
  • N. L. Cassimatis (2003). A Framework for Answering Queries using Multiple Representation and Inference Techniques. In Proceedings of the 10th International Workshop on Knowledge Representation meets Databases. (pdf).