SELF-LEARNING AGENTS: A CONNECTIONIST THEORY OF EMOTION BASED ON CROSSBAR VALUE JUDGMENT

This article presents a computational theory of emotions, based on the theory of emotion as value judgment and appraisal. It assumes that what we feel about other people, events, and things, generally indicates how we evaluate them. The basic assumption is that emotion is an internal self-evaluation of something relevant for the existence of the agent, like self-evaluation of the global state the agent is in, and the behavior the agent is about to perform. This work presents an agent architecture which contains the three components of the control system in biological systems?the genetic, neural, and hormonal component. As distillate of the theory, a working architecture that implements value judgment is presented. The architecture is based on a crossbar connectionist adaptive array, which is designed in a way that it computes from the same crossbar memory elements, both emotions toward encountered situations and emotions toward action tendencies. In such a way it actually builds in hardware, inseparable connections between emotions and behavior. This article gives an instantiation of the architecture and describes a learning experiment to illustrate the emotion learning. A discussion that relates this work to other work reinforcement learning and current research in emotion learning agents is also provided.