A Constrained Mechanism for Procedural Learning

Abstract

The problem studied is that of learning problem solving heuristics by doing. The purpose is to explore the learning behavior of a highly constrained mechanism. The constraints are choosen on the basis of Dsychological considerations. Computer runs show that the constraints do not prevent successful learning. T he learning style of the program turns out to be a function of the structure of the problem space. I HEURISTICS LEARNING The task studied in this paper is that of discovering problem solving heuristics by doing. We presuppose a task-independent, weak problem solver which can take a problem space and a problem as inout, and search for the solution to the latter. We add a set of learning mechanisms which specify how the information generated during search should be encoded for future use. The task of the system is to construct a strategy for searching a problem space on the basis of repeated problem solving trials in that space. Systems of this kind have been described by Anzai & Simon (1979), by Langlev (198?), by Ohlsson (1982; 1983), as well as by others. They will here be called heuristics learners. The purpose of the work reported here is not to construct the most intelligent learning system which the state of the art allows, but to study how a particular set of constraints on an information processing system affects its learning behavior. The constraints chosen for study are based on psychological considerations, although no detailed comparison between the Drogram and human behavior will be made in this paper.