Improving Big Plans

Past research on assessing and improving plans in domains that contain uncertainty has focused on analytic techniques that are exponential in the length of the plan. Little work has been done on choosing from among the many ways in which a plan can be improved. We present the IMPROVE algorithm which simulates the execution of large, probabilistic plans. IMPROVE runs a data mining algorithm on the execution traces to pinpoint defects in the plan that most often lead to plan failure. Finally, IMPROVE applies qualitative reasoning and plan adaptation algorithms to modify the plan to correct these defects. We have tested IMPROVE on plans containing over 250 steps in an evacuation domain, produced by a domain-specific scheduling routine. In these experiments, the modified plans have over a 15% higher probability of achieving their goal than the original plan.