Abstract:
This paper presents a design for a reactive system based on the classical planning techniques of problem reduction and state space search. The proposed approach enables a reactive system to be scaled up to handle larger sets of tasks. Problem reduction synthesizes an initial reactive policy for a given task. When an execution impasse occurs, state space search finds critical choice points, from which control rules are synthesized. These rules alter the policy’s future behavior in order to avoid the impasse. This technique, called "critical choice planning", incrementally debugs the initial policy in the least restrictive way; this "least restrictive property" makes the technique a perfect match for problem reduction. Over time, the problem reduction rules are improved via learning from the debugging experiences.