ABSTRACT
Self-selection of input examples on the basis of performance failure is a powerful bias for learning systems. The definition of what constitutes a learning bias, however, has been typically restricted to bias provided by the input language, hypothesis language, and preference criteria between competing concept hypotheses. But if bias is taken in the broader context as any basis that provides a preference for one concept change over another, then the paradigm of failure-driven processing indeed provides a bias. Bias is exhibited by the selection of examples from an input stream that are examples of failure; successful performance is filtered out. We show that the degrees of freedom are less in failure-driven learning than in success-driven learning and that learning is facilitated because of this constraint. We also broaden the definition of failure, provide a novel taxonomy of failure causes, and illustrate the interaction of both in a multistrategy learning system called Meta-AQUA.
