Acquisition of Appropriate Bias for Inductive Concept Learning”, Thesis Proposal,

I Introduction We consider concept learning problems in which there is a domain of instances over which concepts (generalizations) are to be learned. A trainer presents a sequence of training instances, each labelled as a positive or negative instance of the concept. The task of the learner is to acquire the ability to correctly state, for every instance in the domain, whether that instance is an example of the concept. For any instance which has not been presented by the trainer, the learner must therefore inductively infer whether the instance is an example of the concept. II Problem and Related Work The inductive inference process is driven by two kinds of information. The first kind of information is classifications of training instances given by the trainer. The second kind of information, which we call bias, is broadly defined as anything which influences how the concept learner draws inductive inferences based on the observed training instances. Without such bias, the inductive inference process can perform no better at classifying unobserved instances than random guessing [S]. We are interested here in the issue of how an "appropriate" bias for inductive learning can be acquired automatically