Browse Theses

A learning system is a system that improves its performance with respect to a given task domain over time through its interactions with the task environment. The mechanisms by which such a system manipulates its knowledge about the task environment in response to these interactions constitute the system's "methods of learning". In constructing an artificial learning system, the particular methods employed determine, to a large extent, the ultimate generality of the system. A learning system capable of functioning in a variety of task domains necessarily requires the presence of domain independent methods of learning.

This thesis is concerned with investigating the feasibility of constructing a general purpose learning system around a particular class of domain independent methods called genetic algorithms. To this end, a specific learning system organization, LS-1, is proposed. In further specifying the design, a production system language amenable to manipulation by a genetic algorithm is defined as the system's representation of knowledge, organized as a domain independent framework into which task specific primitives can be injected. The classical genetic algorithms are then modified to suit the specific characteristics of the defined knowledge structure representation. A formal analysis of the search conducted by such a revised genetic algorithm through the space of possible production system programs is performed, demonstrating that it possesses properties analogous to those exhibited by classical genetic algorithms and establishing a sound theoretical foundation for LS-1. Finally, a critic to judge the "relative worth" of a production system program as a potential solution to the task at hand is specified, incorporating both domain independent and task specific sources of judgmental information.

As a demonstration of the feasibility of the design an LS-1 implementation is tested in two distinct and unrelated task domains, each the domain of a related effort in learning system construction. Specifically, the system is faced with (1) a simple maze walk problem and (2) the problem of making the bet decision in draw poker. Initialized in each test with randomly generated production system programs, the LS-1 implementation is shown to rapidly converge on high performance knowledge structures in both task domains, providing empirical evidence of the effectiveness of a genetic algorithm as a general purpose learning mechanism.