Methodological comparison of agent models

Abstract

Hybrid agent architectures comprise the radical change of paradigms in AI over the past decades by reconciling the different styles of reactive, deliberative, even social systems. They have been successfully applied to a range of complex real-world domains. Due to their originally informal background, a verification of design goals in derived implementations, theoretical foundations, and a detailed comparison with other agent models have not yet been obvious. The present work proposes a formal methodology to bridge the gap between theoretical and practical aspects especially of hybrid designs, such as the layered INTERRAP. The employed, connected stages of specification, i.e., architecture, computational model, theory, proof calculus, and implementation, also provide a yet unique framework for comparing heterogeneous agent models including unified and logic-based ones. Based on recent work on INTERRAP, we demonstrate that this methodology allows to compare state-of-the-art designs from robotics, AI, computer science, and cognitive science with respect to a spectrum of inherent properties along the two dimensions of abstraction and declarativity. This supports our claim that INTERRAP is a coherent and advanced account of layered agency including goal-oriented abstraction planning in on-line interaction with reactive skills and social reasoning. We also derive particular research issues to guide the future development of INTERRAP.