Method for evaluating plan recovery strategies in dynamic multi-agent environments

Abstract

ABSTRACT Plan execution in dynamic environments can be affected by unexpected events leading to failures. Research on multi-agent planning area presents recovery strategies with replanning and repairing with evaluation based simply on average values. Thus, in this work, we propose a statistical method to evaluate plan recovery strategies in dynamic environments using a domain-independent approach. To validate the proposed method, we conducted simulated experiments with varying the number of agents, goals, actions, failure probability, and agents’ coupling levels. The evaluation metrics include plan length and planning time. The results highlight with at least 94% certainty that repairing planning time is lower than replanning, and replanning builds plans with fewer actions than repairing. Considering plan recovery strategies in dynamic multi-agent environments, we demonstrate that repairing presents better results as it is faster, but replanning builds better plans as the final plan length is strongly correlated to failure occurrence.