Screening Hypotheses with Explicit Bias

Abstract

This chapter reviews the screening of hypotheses with explicit bias. Bias plays an important role in the major subfield of machine learning called empirical induction. Explicit bias offers advantages in addition to helping a system select generalization operators. Empirical induction of a concept from examples is performed by formulating hypotheses that approximate an unknown target concept based on experience with instances classified as positive or nega­tive examples of the target concept. If learning is incremental, hypotheses are formed and then modified by generalization or specialization to remain consistent with a growing set of known instances. Three major techniques exist for biasing empirical induction. The first technique consists of using a res­tricted hypothesis language. The second bias technique is testing. The third technique called screening consists of selecting a generalization operator rather than generating the alternative hypotheses. This latter technique is less widely used than the other two but offers advantages when the bias is explicit. PREDICTOR is a system that uses explicit bias in its learning method. PREDICTOR focuses on the screening method of bias.