Symbolic Stochastic Focused Dynamic Programming with Decision Diagrams

We present a stochastic planner based on Markov Decision Processes (MDPs) that participates to the probabilistic planning track of the 2006 International Planning Competition. The planner transforms the PPDDL problems into factored MDPs that are then solved with a structured modified value iteration algorithm based on the safest stochastic path computation from the initial states to the goal states. First, a state subspace is computed by making all the transitions deterministic. Then, a step of modified value iteration on the current reachable state subspace alternates with a step of reachable state expansion by following the current policy.