Predicting and Explaining Success and Task Duration in the Phoenix Planner*

Phoenix is a multi-agent planning system that fights simulated forest fires. In this paper we describe an experiment with Phoenix in which we uncover factors that affect the planner's behavior and test predictions about the planner's robustness against variations in some of these factors. We also introduct a technique – path analysis – for constructing and testing causal explanations of the planner's behavior. * This research was supported by DARPA Contract F49620-C-89-00113; by an Intelligent Real-time Problem Solving Initiative Contract AFOSR-91-0067; by NSF under an Issues in Real-time Computing grant, CDA-8922572; and by Texas Instruments Corporation. Predicting & Explaining Success in Phoenix Hart and Cohen