Edmisten@SUMEX-AIM.ARPA (09/25/84)
From: Paula Edmisten <Edmisten@SUMEX-AIM.ARPA>
[Forwarded from the Stanford SIGLUNCH distribution by Laws@SRI-AI.]
DATE: Friday, September 28, 1984
LOCATION: Chemistry Gazebo, between Physical and Organic Chemistry
TIME: 12:05
SPEAKER: John Laird,
Xerox Corp.
ABSTRACT: SOAR: An Architecture for General Intelligence
I will present recent progress in developing an architecture for general
intelligence, called Soar. In Soar, all problem solving occurs as
search in a problem space and all knowledge is encoded as production
rules. I will describe the Soar architecture and then present three
demonstrations of its generality and power.
1. Universal Subgoaling: All subgoals are created automatically by the
architecture whenever the problem solver is unable to carry out the
basic functions of problem solving (so that all subgoals in Soar are
also meta-goals). All the power of Soar is available in the subgoals,
including creating new subgoals, making Soar a completely reflective
problem solver.
2. A Universal Weak Method: The weak methods emerge from knowledge about
a task instead of through explicit representation and selection.
3. R1-Soar: Although Soar was designed for general problem-solving, it
is also effective in the knowledge-intensive domains of expert systems.
This will be demonstrated by a partial implementation of the R1 expert
system in Soar.
Soar also has a general learning mechanism, called Chunking. Paul
Rosenbloom will present this aspect of our work at the SIGLunch on
October 5.