An Empirical Analysis of Local Search in Stochastic Optimization for Planner Strategy Selection

Optimization of expected values in a stochastic domain is common in real world applications. However, it is often difficult to solve such optimization problems without significant knowledge about the surface defined by the stochastic function. I n this paper we examine local search techniques to solve stochastic Optimization. I n particular, we analyze assumptions of smoothness upon which these approaches often rely. We examine these assumptions i n the context of optimizing search heuristics for a plannerlscheduler on two problem domains. We compare three search algorithms to improve the heuristic sets and show that these algorithms work because their local steps i n our domains have relevant smoothness properties.