Bounded Suboptimal Game Tree Search

Abstract

Finding the minimax value of a game is an important problem in a variety of fields, including game theory, decision theory, statistics, philosophy, economics, robotics, and security. Classical algorithms such as the Minimax algorithm can be used to find the minimax value, but require iterating over the entire game tree, which is in many cases too large. Alpha-Beta pruning identifies portions of the game tree that are not necessary for finding the minimax value, but in many cases the remaining part of the game tree is still too large to search in reasonable time. For such cases, we propose a class of algorithms that accepts a parameter e and returns a value that is guaranteed to be at most e away from the true minimax value. We lay the theoretical foundation for building such algorithms and present one such algorithm based on Alpha-Beta. Experimentally, we show that our algorithm allows controlling this runtime/solution quality tradeoff effectively.