Pessimistic Heuristics Beat Optimistic Ones in Real-Time Search
Aleksander Sadikov, Ivan Bratko · European Conference on Artificial Intelligence · 2006
Admissibility is a desired property of heuristic evaluation functions, because when these heuristics are used with complete search methods, such as A* and RBFS, they guarantee that an optimal solution will be found. Since every optimistic heuristic function is admissible, optimistic functions are widely used. We show, however, that with incomplete, real-time search, optimistic functions lose their appeal, and in fact they may hinder the search under quite reasonable conditions. Under these conditions the exact opposite is to be preferred, i.e. pessimistic heuristic functions that never underestimate the difficulty of the problem. We demonstrate that such heuristics behave better than optimistic ones of equal quality on a standard testbed using RTA* search method.