Intelligent Systems And Their Societies Walter Fritz

SOAR

 

Soar was created by John Laird, Allen Newell (of Carnegie Mellon University),and Paul Rosenbloom. This program was described in the SIGART journal of August '91 by John Laird, Mike Hucka, and Scott Huffman of the Artificial Intelligence Laboratory of the University of Michigan and by Paul Rosenbloom of the Information Sciences Institute of the University of Southern California. It was started in1983 and is still under development, the most recent version (2003) is Soar 8.2. It is being used in many countries around the world.

This IS includes the creation of sub objectives, learning and interaction with an external environment. Response rules, (called productions), are grouped into "problem spaces", meaning those useful for a specific problem. When needed, those of one problem space can call those of another. For instance, there is a selection problem space, one for interaction with the environment and another for the arithmetic problem space.

Response rules are learned and they have composite responses, called "operators". SOAR uses sub objectives when it cannot find an applicable response rule. Once the sub objective has been satisfied, all related responses are incorporated into a new response rule with a composite response, so that in the future this can be used directly instead of having to use sub objectives.

 

SOAR learns by itself. The tasks of acquisition of information and the decomposition of general responses into specific ones function in parallel. The activities of comparison, decision, and learning are basic mechanisms; the program cannot modify them. It is noteworthy that the authors have combined the functions of creating the situation and selecting a response rule into one function, called "working memory". This is a type of blackboard that receives sense information, in it situations and plans are built up and from it the motor modules receive the commands they are to execute.

SOAR's designers applied the system to problems in many fields such as solving simple puzzles and playing simple games, medical diagnosis, the control of a robot, and others. In 1991 the system was learning human languages and typically learned about 1000 response rules and reacted in about 25 milliseconds.

 

For continuous reading, like a book - continuehere
Jump to the book Index / Artificial IS /
Proposed brain / top of the SOAR.


Last Edited 15 Sept 2011 / Walter Fritz
Copyright © New Horizons Press