2 nd European Workshop on Planning ( EWSP ’ 93 ) Using Goals to Find Planswith High Expected

We describe a method for planning to achieve goals under uncertainty that makes use of decision-theoretic methods to guide search. Given a probabilistic model of the world and a utility measure on world states, we wish to nd plans (sequences of actions) with high expected utility. Finding a plan maximizing the expected utility is combinatorial in nature. In previous research, we coped with the combinatorics by making simplifying assumptions that sometimes led to a poor choice of plan. In this paper, we reduce the combinatorics by restricting attention to plans that are likely to achieve speciic goals; we then use a successive approximation algorithm to select from these plans one with high utility. We obtain the restricted set of plans using a procedure that can, given a goal, generate candidate plans one at a time as needed; these plans are produced in decreasing order of probability of achieving their goal. This procedure is also used to obtain iteratively reened bounds on the expected utility of the candidate plans; these bounds are used in choosing a plan to be executed, hence our method produces better plans the more time it is given. We apply this method to a robotics problem addressed in our previous research and show that the goal-directed search produces better plans.