Unifying control in a layered agent architecture

Abstract

In this paper, we set up a unifying perspective of the individual control layers of the architecture InteRRaP for autonomous interacting agents. InteRRaP is a pragmatic approach to designing complex dynamic agent societies, e.g. for robotics Müller & Pischel and cooperative scheduling applications Fischer et al.94. It is based on three general functions describing how the actions an agent commits to are derived from its perception and from its mental model: belief revision and abstraction, situation recognition and goal activation, and planning and scheduling. It is argued that each InteRRaP control layer - the behaviour-based layer, the local planning layer, and the cooperative planning layer - can be described by a combination of different instantiations of these control functions. The basic structure of a control layer is defined. The individual functions and their implementation in the different layers are outlined. We demonstrate various options for the design of interacting agents within this framework by means of an interacting robots application. The performance of different agent types in a multiagent environment is empirically evaluated by a series of experiments.