R* Search

Abstract

Optimal heuristic searches such as A* search are widely used for planning but can rarely scale to large complex problems. The suboptimal versions of heuristic searches such as weighted A* search can often scale to much larger planning problems by trading off the quality of the solution for efficiency. They do so by relying more on the ability of the heuristic function to guide them well towards the goal. For complex planning problems, however, the heuristic function may often guide the search into a large local minimum and make the search examine most of the states in the minimum before proceeding. In this paper, we propose a novel heuristic search, called R* search, which depends much less on the quality of the heuristic function. The search avoids local minima by solving the whole planning problem with a series of short-range and easy-to-solve searches, each guided by the heuristic function towards a randomly chosen goal. In addition, R* scales much better in terms of memory because it can discard a search state-space after each of its searches. On the theoretical side, we derive probabilistic guarantees on the sub-optimality of the solution returned by R*. On the experimental side, we show that R* can scale to large complex problems.