Planning with deadlines in stochastic domains

摘要: We provide a method, based on the theory of Markov decision problems, for efficient planning in stochastic domains. Goals are encoded as reward functions, expressing desirability each world state; planner must find policy (mapping from states to actions) that maximizes future rewards. Standard goals achievement, well maintenance and prioritized combinations goals, can be specified this way. An optimal found using existing methods, but these methods at best polynomial number domain, where is exponential propositions (or state variables). By information about starting state, function, transition probabilities we restrict planner's attention set likely encountered satisfying goal. Furthermore, generate more or less complete plans depending time it has available. describe experiments involving mobile robotics application consider problem schedulilng different phases algorithm given constraints.