Pathology in heuristic search

Most game-playing programs use minimax search: they build a game tree, heuristically evaluate the leaves, back up the heuristic values to the root and use them to select a move. Practice shows that the deeper such programs search, the better they play. However, mathematical models intended to explain this formally paradoxically showed that deeper search results in worse play [6,1]. This phenomenon was termed minimax pathology. The most common explanation for the pathology is that the mathematical models lacked dependence between nearby positions, which is clearly present in games. In the Ph.D. thesis we propose an explanation that does not require dependence and uses real-number position values. Since gameplaying programs use discrete values, we also investigate the effect of the number of possible values as well as other parameters on the pathology. Finally we return to the question that started pathology research: why is minimax search beneficial. Pathology was recently observed in single-agent search as well [2]. We investigate the conditions under which it appears and show that it has a practical impact by observing it in path-finding on maps from commercial computer games. 2. Pathology in minimax search