Wolverton, M. and Washington, R. Segmenting Reactions to Improve the Behavior of a Planning/Reacting Agent, in Proceedings of the Third International Conference on Artificial Intelligence Planning Systems (AIPS-96), May 1996.
An agent operating in a real-world environment will inevitably encounter some events that demand very quick response—the deadline for an action is short, and the consequences of not acting are high. For these types of events, the agent has a better chance of acting appropriately if it has a pre-stored set of reactions that can be quickly retrieved based on features of the situation. In this paper, we examine the problem of selecting the best set of reactions for an agent to store. In particular, we examine the benefit of including *intervals* of reactions—i.e., executing some reactions only across segments of the complete numeric ranges over which they are defined. We present a decision-theoretic algorithm for selecting the optimal set of reaction intervals, and we present experiments with a computer implementation of that algorithm, called KNEEJERK. The experiments show that the benefit of breaking down reactions into intervals is quite high under a wide range of circumstances.