Layered Concept-Learning and Dynamically Variable Bias Management

Abstract

Concept learning is inherently complex. Without severe constraint or inductive "bias," the general problem is intractable. While most learning systems have been designed with built-in biases, these systems typically work well only in narrowly circumscribed problem domains. Here we present a model of concept formation that views learning as a simultaneous optimization problem at three different levels, with dynamically chosen biases guiding the search for satisfactory hypotheses. In this model, the partitioning of events into classes occurs through dynamic interactions among three layers: event space, hypothesis space, and bias space. This view of the induction process may help clarify the problem of learning and lead to more general and efficient induction systems. To test this model of meta- knowledge, a variable bias management system (VBMS) has been designed and partly implemented. The system wiU dynamically alter evolving hypotheses, concept representation languages, and concept formation algorithms by monitoring progress and selecting biases based on characteristics of the particular induction problems presented. VBMS is designed to learn the best biases for different types of induction problems. Thus it is robust (effective and efficient in many domains). The system can learn incrementally despite noisy data at any level.