Partial Satisfaction (Over-Subscription) Planning as Heuristic Search

Many planning problems can be characterized as over-subscription problems in that goals have different utilities, actions have different costs and the planning system must choose a subset that will provide the greatest net benefit. This type of problems can not be solved by existing planning systems, where goals are assumed to have uniform utility, and the planner can terminate only when all of the goals are achieved. Existing methods for such problems use greedy approaches, which pre-select a subset of goals based on their estimated utility, and solve for those goals. Unfortunately, greedy methods, while efficient, can produce plans of arbitrarily low quality. In this paper, we introduce a more sophisticated heuristic search framework for over-subscription planning problems. In our framework, top-level goals are treated as soft-constraints and the search is guided by a relaxed-plan based heuristic that estimates the most beneficial set of goals from a given state. We implement this search framework in the context of Sapa, a forward state-space planner. We provide preliminary empirical results that demonstrate the effectiveness of our approach in comparison to a greedy approach.