Repairing decision-theoretic policies using goal-oriented planning

In this paper we address the problem of how decision-theoretic policies can be repaired. This work is moti- vated by observations made in robotic soccer where decision- theoretic policies become invalid due to small deviations dur- ing execution; and plan repair might pay off rather than re- plan from scratch. Our policies are generated with Readylog , a derivative of Golog based on the situation calcu- lus, which combines programming and planning for agents in dynamic domains. When an invalid policy was detected, the world state is transformed into a PDDL description and a state-of-the-art PDDL planner is deployed to calculate the repair plan.