In computational studies of intelligent behavior several phenomena have been isolated, among the most important of which are learning, reasoning, and planning. Obstacles that have been found pervasive are that the formalization of the various phenomena are often inconsistent with each other, and, even more often, that they are computationally intractable. This main thrust of this research project is to develop a computational framework and some algorithms that would comprise a unified approach to this problem area. The proposed starting point is the idea that the concept of probably approximately correct (PAC) learning provides a quantitative criterion on the behavioral performance of a system. These analytic methods used in computational learning theory, are extended to a wider class of phenomena, including various notions of reasoning and planning.