Controlling an autonomous agent using internal value based action selection

Abstract

In this paper we describe an approach of controlling an autonomous robot by means of a hierarchical control structure, with a learning action selection. Since Damasio's "Descartes' error" in 1994 the number of approaches to action selection that use internal values, derived from psychological models of emotions or drives has increased significantly. The approach realises a learning action selection mechanism in a hierarchy of sensory and actuatory layers. The sensory values yield the internal states, as a basis for action selection. In addition they are used to calculate the reinforcement signal that trains the action selection.