Theory and practice of hybrid agents

Abstract

Hybrid Agents integrate different styles of reactive, deliberative, and cooperative problem solving in a modular fashion. They are the prime device of (Distributed) Artificial Intelligence and Cognitive Science for realising a broad spectrum of simultaneous functionalities in application domains such as Artificial Life, (Tele-)Robotics, Flexible Manufacturing, and Automated Transportation. In this report we propose a design methodology for hybrid agents which combines complementary approaches of Software Engineering and declarative Cognitive Robotics at five interconnected specification stages: Architecture, Computational Model, Theory, Inference, and Implementation. The design methodology is then applied to the reconstruction of the layered agent model InteRRaP ('Integration of Reactivity and RAtional Planning'). InteRRaP, in spite of its practical success in the past, suffers from its originally architecture-centred and informal description. The result is an agent model, InteRRaP-R (the additional 'R' stands for 'Resource-adapting'), which maps its layered architecture onto the formally described interplay of concurrent processes whose runtime is scheduled by meta-control: The processes of a higher layer control the processes of the subordinate layer by the allocation of (computational) resources. Processes are continuous computational activities that are realised as situated inference procedures which implement well-defined subsets of a common logic of time and action. Three representative scenarios are chosen for an evaluation of InteRRaP-R: the Automated Loading Dock, the RoboCup Simulation League, and the ROTEX space robot. These case studies confirm the applicability of our slogan "Agent = Logic + Architecture" to the theory and practice of hybrid agents.