HINTON@CMU-CS-C.ARPA (04/19/84)
From: Geoff Hinton <HINTON@CMU-CS-C.ARPA> [Forwarded from the CMU-AI bboard by Laws@SRI-AI.] "RECENT PROGRESS IN SOAR: CHUNKING AND R1-SOAR" by John Laird & Paul Rosenbloom AI Seminar, Tuesday April 24, 4.00pm, Room 5409 In this talk we present recent progress in the development of the @p[Soar] problem-solving architecture as a general cognitive architecture. This work consists of first steps toward: (1) an architecture that can learn about all aspects of its own behavior (by extending chunking to be a general learning mechanism for @p[Soar]); and (2) demonstrating that @p[Soar] is (more than) adequate as a basis for knowledge-intensive (expert systems) programs. Until now chunking has been a mechanism that could speed up simple psychological tasks, providing a model of how people improve their performance via practice. By combining chunking with @p[Soar], we show how chunking can do the same for AI tasks such as the Eight Puzzle, Tic-Tac-Toe, and a portion of an expert system. More importantly, we present partial demonstrations: (1) that chunking can lead to more complex forms of learning, such as the transfer of learned behavior (that is, the learning of generalized information), and strategy acquisition; and (2) that it is possible to build a general problem solver that can learn about all aspects of its own behavior. Knowlege-intensive programs are built in @p[Soar] by representing basic task knowledge as problem spaces, with expertise showing up as rules that guide complex problem-space searches and substitute for expensive problem-space operators. Implementing a knowledge-intensive system within @p[Soar] begins to show how: (1) a general problem-solving architecture can work at the knowledge intensive (expert system) end of the problem solving spectrum; (2) it can integrate basic reasoning and expertise, using both search and knowledge when relevant; and (3) it can perform knowledge acquisition by transforming computationally intensive problem solving into efficient expertise-level rules (via chunking). This approach is demonstrated on a portion of the expert system @p[R1], which configures computers.