Generating Plans To Succeed in Uncertain Environments

Planners have traditionally not handled domain uncertainty, postponing that possibility to error monitoring routines during the execution of the plan. In real-world domains with incomplete knowledge, this results in inevitable delays due to replanning. This paper describes a planner that considers the reliability of the agent's actions (learned from previous experience) while generating a plan. This is done by incorporating into the domain representation, the probabilities that the effects of an action will be observed after its execution. These probabilities may depend on the current state of the environment, allowing the formation of hard and soft constraints for actions. Action selection is performed by computing an "expected utility" for each action by a bidirectional spreading activation process which propagates goal utilities backward and predicted states of the environment forward. This connectionist approach allows the simultaneous generation of multiple plans, resulting in the availability of fall-back plans if the one with the highest probability of succeeding fails.