Learning to be Competent

Abstract

The thesis presents a new approach for the study of competent cognitive behavior. The approach, learning to be competent, suggests that learning phenomena and the competencies attributed to intelligence should be studied together. Instead of requiring omniscience or otherwise optimal performance, we claim that the tasks and success criteria should be defined behaviorally; that is, a system is competent if it functions well in its environment. We further suggest that competent behavior should only be expected in light of a learning experience in the same or similar environment, and that the solutions exhibited should be computationally efficient. These ideas are presented in a formal setting, so that the various tasks and their proposed solutions can be studied and analyzed. Thus, one contribution of this approach is in formalizing the problem in a form that is amenable to analysis, while at the same time being cognitively and computationally plausible. The learning to reason framework is used to study the problem of logical reasoning in propositional domains. We consider a variety of possible interfaces for learning, and describe learning algorithms that interact with them, thus demonstrating the robustness of this approach. The results show that learning to reason is possible even in cases where the traditionally separate problems, namely concept learning and reasoning by proving assertions, do not have efficient solutions. In the course of studying reasoning tasks, we develop a model based representation, the set of characteristic models, which supports efficient solutions for several forms of logical reasoning. This representation is utilized in the learning to reason framework, and helps in deriving the above mentioned results. The representation is also shown to have other applications, in the theory of relational databases, and in computational tasks that arise in the design of such databases. The task of acting in a dynamic world in order to achieve some goals is studied in the learning to act framework. We present results on supervised learning of action strategies in the form of production rule systems. The framework and the results combine features from the area of symbolic computation and that of reactive agents, which have been previously seen as opposed if not contradictory, and thus advance our understanding of the problems.