Failure Detection and Reactive Teaming for Behavior-Based Subsumption

Abstract

Fast, lightweight intelligent controllers make it possible to apply agent-based artificial intelligence to even highly resource-constrained systems. Reactive control methods can provide this capability, and are frequently used in applications such as real-time games and robotics. Unfortunately, reactive controllers may not deal well with errors, as they cannot fall back on a planner to deal with unexpected situations. This limitation makes them more susceptible to certain types of failures than deliberative techniques. In this paper, we describe how four major types of failures can be detected in reactive controllers. We show how this can lead to extensions for behavior-based subsumption which allow more robust single agents, enable inexpensive multi-agent coordination, and improve agent design tools.