Naval Research Laboratory, Code 5514 Navy Center for Applied Research in Artificial Intelligence 4555 Overlook Ave., S.W., Washington, D.C. 20375-5320

Reinforcement Learning Methods (RLMs) typically select candidate solutions stochastically based on a credibility space of hypotheses which the RLM maintains, either implicitly or explicitly. RLMs typically have both inductive and deductive aspects: they inductively improve their credibility space on a stage-by stage basis; they deductively select an appropriate response to incoming stimuli using their credibility space. In this sense, RLMs share some learning attributes in common with active, incremental concept learners. Unlike some concept learners that employ deterministic procedures for selecting hypotheses, however, the evaluations of hypotheses provided to RLMs are often uncertain, either due to noisy environments, or due to summary evaluations which occur after a sequence of learner-environment interactions. This paper examines issues of inductive learning bias in this context experimentally. Specifically, the paper addresses inductive learning biases in the context of a simple RLM called a Collective Learning Automaton (CLA). The CLA learns the shortest path through a small network. The research points out some of the difficulties of finding performance measures that indicate the strongest, correct biases for the automaton.