Using a Plan Graph with Interaction Estimates for Probabilistic Planning

Many planning and scheduling applications require the ability to deal with uncertainty. Often this uncertainty can be characterized in terms of probability distributions on the initial conditions and on the outcomes of actions. These distributions can be used to guide a planner towards the most likely plan for achieving the goals. This work is focused on developing domain-independent heuristics for probabilistic planning based on this information. The approach is to first search for a low cost deterministic plan using a classical planner. A novel plan graph cost heuristic is used to guide the search towards high probability plans. The resulting plans can be used in a system that handles unexpected outcomes by runtime replanning. The plans can also be incrementally augmented with contingency branches for the most critical action outcomes.