by Patrick Henry Winston
Professor and Former Director, Artificial Intelligence Laboratory, MIT
Past President, American Association of Artificial Intelligence
Soar is a candidate unified theory of cognition embodied in a computational programming architecture, developed by John Laird, Paul Rosenbloom, and Allen Newell starting in 1982 at Carnegie Mellon University. Its development continues through the efforts of scientists worldwide in the areas of artificial intelligence, cognitive science and human-computer interaction.
Soar uniformly represents short-term knowledge as a network of active symbols. Long-term knowledge is a set of condition-action rules. The conditions of each rule form a pattern to match against the active symbol network. When a rule's condition matches, the rule executes by performing its actions. These actions may result in adding (or deleting) symbols in the short-term knowledge structure.
To manage complexity, Soar includes a goal hierarchy, allowing successive decomposition of problems into component subproblems. Soar includes mechanisms to create new goals automatically in response to a system's long-term knowledge and current situation.
Unlike conventional programming languages, Soar does not enforce a serial flow of control. Rather, actions occur any time an associated pattern matches (via a condition-action rule). Multiple rules may fire in parallel, and Soar provides preference mechanisms and automatic subgoals to handle conflicts, if they occur. Soar represents perceptual and conceptual knowledge uniformly in short-term memory. Thus, new actions flow from previous actions and from changes in the external environment. This combination allows Soar systems to engage in interrupt-driven behavior in the same manner as they direct action toward explicit goals. The rule system incorporates the latest pattern-matching technology, allowing rapid processing. Thus Soar is extremely well suited for the development of intelligent systems that must generate actions in time comparable to human decision time.
Soar includes an automatic learning mechanism inspired by the psychological concept of chunking. Soar compiles sequences of actions into new units of knowledge (chunks) that can 'short-circuit' some reasoning steps when the system faces similar situations in the future. New chunks fit uniformly into a system's existing long-term rule set. Thus, a Soar system can incrementally learn new facts about the world, as well as more efficient representations of its initial long-term knowledge.
Since its development, the Soar architecture has been used to develop a variety of research systems and commercial applications. Examples include NL-Soar (a Soar sub-system that allows programs to learn and interpret natural English text), TacAir-Soar (a synthetic combat pilot, which performs military air missions in real-time distributed simulation environments), and KB-agent (a system for automating business claims and high volume processing).
 |
|
Photo by: Photographer’s Mate 2nd Class Michael |
©2008 Soar Technology, Inc. All rights reserved
Legal Notice | Site Map