Strategy Acquisition in a General Architecture for Intelligence

Objectives

In this project, we plan to re-implement a proven cognitive model for strategy acquisition in a more general artificial intelligence architecture. Specifically, we will re-implement Jones and Van Lehn's (1994) GIPS system in Soar (Laird et. al., 1987), replacing the probabilistic learning mechanism with Miller's (1993) SCA. SCA is a concept acquisition system that has successfully been used to model other human learning phenomena. In doing so we intend to (1) fill in process details for the cognitive model where GIPS relies on a descriptive model, and (2) explore the power of SCA as a general learning subsystem in a larger context. We feel that combining the high-level problem solving capabilities of GIPS with the lower-level learning mechanisms of SCA will provide a "soup to nuts" process model for strategy acquisition, and will serve as a springboard for future research.

Background

Siegler and Jenkins (1989) provide a detailed analysis of the individual behaviors of several children, each of whom is just learning to add small numbers. They focus changes which occur over time to the procedure that each child uses to perform addition. Over time, the children seemed to discover a series of more efficient methods for computing the sum of two addends, and this series seemed to follow a fairly regular progression. Siegler and Jenkins termed this process of exploring alternative procedures strategy acquisition, defining a strategy to be a goal directed but non-compulsive procedure; in other words, a strategy describes a method by which a task can be accomplished, but this method may be one of several exhibited by the learner under different circumstances. They note that the progression of behaviors exhibited by the children is not well described by current computational learning theory; specifically, the learning does not seem to be a case of simple speedup, nor does it seem to be driven by problem-solving impasses in the normal sense.

Rising to the challenge, Jones & Van Lehn proposed a system called GIPS (General Inductive Problem Solver), which provides a process model of the phenomena that Siegler and Jenkins observed. At GIPS' heart is a means-ends analysis performance system that learns operator selection criteria with a probabilistic mechanism under supervised learning conditions. The system is remarkable in that it is able to accurately model the aggregate sequence of strategies through which most children seem to normally progress, and does so with a tight set of independently-motivated mechanisms.

Implementation Plan & Schedule

For purposes of this project, we will apply the system to a knowledge-lean domain where we can observe acquisition of means-ends behavior in very controlled situations. The implementation of the system falls naturally into several phases:

• Embed SCA into a performance system that allows it to make operator selection decisions. This phase has successfully been completed for the infamous monkey-and-bananas domain.
• Implement a limited learning mechanism that allows SCA to acquire operator selection rules as in GIPS. In this phase, the learning will be much more supervised than in GIPS; specifically, the system will be trained at every choice point. This phase is currently underway. Our deliverable here is a system that can successfully train and perform under rigid supervision in the task environment. Our target for completion is March 15.
• Determine appropriate boundaries between "cognitive" and "perceptual/motor" systems. Although unimportant in the original implementation of GIPS, these issues become important in any implementation that uses the Soar architecture. This phase may be intertwined with the previous phase. Our deliverable here is a system that clearly separates the internal problem solving from the external perception and action systems. This version will have a graphical interface that at minimum displays the actions which the system is taking. Our target for completion is March 31.
• Implement the GIPS credit assignment mechanism. GIPS credit assignment mechanism is more sophisticated than the mechanism described above. Specifically, it uses chronological backtracking to search for a solution, maintaining the current path and each operator that led off the path. Once a goal state is reached, it traces back through the solution path, training each operator on the path as a positive training instance and training each operator that led off the path as a negative instance. Our target for completion is April 23.

Plans for Future Work

After completing the core system, we plan to model the children's addition domain. We will begin with GIPS' implementation of the addition domain, possibly enriching it with (1) a credit assignment mechanism that allows the system to learn with little or no supervision, (2) a distinct model of the processes that are "external" to the cognitive component of the system, namely, the perceptual and motor components, and (3) a more detailed theory of the child's concept of number. We hope to parameterize our system in such a way that we will be able to account for the individual protocols from subjects in Siegler & Jenkins' experimental data.

References

R. Jones & K. Van Lehn. "Acquisition of Children's Addition Strategies: A Model of Impasse- Free, Knowledge-Level Learning," Machine Learning, 16:11-36. 1994.

J. E. Laird, A. Newell, and P. Rosenbloom. "Soar: An Architecture for General Intelligence," Artificial Intelligence, 33:1-64. 1987.

C. Miller. "Modeling Concept Acquisition in the Context of a Unified Theory of Cognition," Ph.D. Thesis. CSE Technical Report 157-93, The University of Michigan. 1993.

R. Siegler & E. Jenkins. How Children Acquire New Strategies. Lawrence Erlbaum Associates. Hillsdale, New Jersey. 1989.

Last Modified: March 10, 1996