Computer Science Department
School of Computer Science, Carnegie Mellon University
Chunking in Soar: The Anatomy of a General Learning Mechanism
John E. Laird oot Intelligent Systems Laboratory, Xerox Palo Alto Research Center , Paul S. Rosenbloom Departments of Computer Science and Psychology, Stanford University , Allen Newell
The goal of the Soar project is to build a system capable of general intelligent behavior. We seek to understand what mechanisms are necessary for intelligent behavior, whether they are adequate for a wide range of tasks - including search-intensive tasks, knowledge-intensive tasks, and algorithmic tasks - and how they work together to form a general cognitive architecture. One necessary component of such an architecture, and the one on which we focus in this paper, is a general learning mechanism. Intuitively, a general learning mechanism should be capable of learning all that needs to be learned. To be a bit more precise, assume that we have a general performance system capable of solving any problem in a broad set of domains. Then, a general learning mechanism for that performance system would possess the following three properties.
Task generality . It can improve the system's performance on all of the tasks in the domains. The scope of the learning component should be the same as that of the performance component.
Knowledge generality . It can base its improvements on any knowledge available about the domain. This knowledge can be in the form of examples, instructions, hints, its own experience, etc.
Aspect generality . It can improve all aspects of the system. Otherwise there would be a wandering-bottleneck problem oot Mitchell, T.M., Learning and Problem Solving, In Proceedings of IJCAI-83 . Los Gatos, CA:Kaufmann, 1983. ,in which those aspects not open to improvement would come to dominate the overall performance effort of the system.
These properties relate to the scope of the learning, but they say nothing concerning the generality and effectiveness of what is learned. Therefore we add a fourth property.
Transfer of learning. What is learned in one situation will be used in other situations to improve performance. It is through the transfer of learned material that generalization , as it is usually studied in artificial intelligence, reveals itself in a learning problem solver.
Generality thus plays two roles in a general learning mechanism: in the scope of application of the mechanism and the generality of what it learns.
There are many possible organizations for a general learning mechanism, each with different behavior and implications. Some of the possibilities that have been investigated within AI and psychology include:
A Multistrategy Learner. Given the wide variety of learning mechanisms currently being investigated in AI and psychology, one obvious way to achieve a general learner is to build a system containing a combination of these mechanisms.
A Deliberate Learner. Given the breadth required of a general learning mechanism, a natural way to build one is as a problem solver that deliberately devises modifications that will improve performance. The modifications are usually based on analyses of the tasks to be accomplished, the structure of the problem solver, and the system's performance on the tasks. Sometimes this problem solving is done by the performance system itself, or in a production system that employs a build operation--whereby productions can themselves create new productions. Sometimes the learner is constructed as a separate critic with its own problem solver, or as a set of critics.
A Simple Experience Learner. There is a single learning mechanism that bases its modifications on the experience of the problem solver. The learning mechanism is fixed, and does not perform any complex problem solving.
The third approach, the simple experience learner, is the one adopted in Soar .