Control strategies for a stochastic planner

摘要: We present new algorithms for local planning over Markov decision processes. The base-level algorithm possesses several interesting features control of computation, based on selecting computations according to their expected benefit quality. are shown expand the agent's knowledge where world warrants it, with appropriate responsiveness time pressure and randomness. then develop an introspective algorithm, using internal representation what computational work has already been done. This strategy extends base warranted by model put into various parts this model. It also enables agent act so as take advantage savings inherent in staying known state space. flexibility provided strategy, incorporating natural problem-solving methods, directs effort towards it's needed better than previous approaches, providing greater hopes scalability large domains.