Adaptive Agent Design Based on Reinforcement Learning and Tracking.pdfVIP

Adaptive Agent Design Based on Reinforcement Learning and Tracking Alfred D.M. Wan and Peter J. Braspenning MATRIKS, Dept. of Computer Science, University of Limburg, P.O. Box 616, Maastricht, The Netherlands e-mail: [wan,braspenning]@cs.rulimburg.nl Abstract AI seems to be developing toward a theory of intelligent agents (IA), however, the theoretical basis which should underly the notion of agency has remained strikingly underdeveloped. This paper is intended to improve on this situation, by analysing the most basic notions necessary for a theory of agency: action, self, control, autonomy and adaptivity. Our main thesis is that autonomy is the essential concept in agency and that autonomy implies adaptivity. We supply a description of an adaptive autonomous agent (A3), based on the outlines given by Ashby [3]. Furthermore, we indicate how the two-folded adaptive powers, i.e. regulation and learning can be merged into one Rein- forcement Learning (RL) system. We describe our RL implementation, which solves the as yet unexplored problem of dynamic goal state attainment. 1 Introduction Recently AI seems to be developing toward a science of intelligent agents (IA), rather than toward a comprehensive theory of intelligence1 . By no means, however, does a (remotely mature) theory of agency or agenthood exist, whereas theories of multi-agent interactive communication and organizational behaviour are lacking too. Some authors endorse a very open view on the definition of what an agent is, indeed. One of the most extreme (but certainly