"Learning to be Competent"at the Department of Computer Science, University of Helsinki.
Date Friday, October 18th, 1996 Place Room A516 (in the 5th floor, next to the elevators) Time 12:15 (The lecture will last approximately an hour)
Learning to be Competent is a general approach to the study of cognitive competent behavior. The main tenet of the approach is that learning and competence modules of intelligent agents should and can be studied together. The framework requires that plausible learning interfaces, and the tasks or competencies be defined, and suggests that the performance of an agent be measured in behavioral terms. That is, instead of requiring omniscience or otherwise optimal behavior, we say that an agent is competent if it functions well in its environment. These ideas are formalized in the framework of learning to be competent. Using this formalization, new algorithms are derived and their performance analyzed, so as to show that the combined task of learning a particular competence is tractable.My talk will first discuss the general features of the approach and then review some of the results that have been achieved. In particular, the competence of logical reasoning in propositional domains is formalized as the problem of learning to reason. The task of acting in a dynamic stochastic world in order to achieve some goals is formalized as the problem of learning to act. In both cases we show that learning to be competent is possible even when the traditionally separate problems, namely learning a description of the world, and performing the appropriate competence using this description, do not have efficient solutions. Some of the results reviewed in the talk are joint work with Dan Roth.