Using Proofs and Refutations to Learn from Experience

Abstract

To learn, a learner needs to formulate plans, monitor the plan execution to detect violated expectations, and then diagnose and rectify errors which the dis-confirming data reveal. In this paper, five heuristic methods are presented for repairing flawed beliefs. These beliefs are considered as theories that predict effects of actions. Theories presuppose particular structural characteristics. When data disconfirm a theory, the heuristics proposed suggest specific ways to remedy the theory, including restricting the conditions for invoking the theory and weakening the theory’s predictions. The five methods accomplish retraction, exclusion, avoidance, assurance and inclusion of outcomes that disconfirm a theory’s predictions. Each proposed theory fix produces as a by-product new domain concepts that capture environmental characteristics of instrumental value to the learner. The techniques proposed here provide the first analytical methods for constructing new knowledge. They extend and make practical the ideas of proofs and refutations originally introduced by Lakatos.