Local versus global tables in Minimax Game Search

Abstract

Minimax Game Search with alpha-beta pruning can utilize heuristic tables in order to prune more branches and achieve better performance. The tables can be implemented using different memory models: global tables, worker-local tables and processor-local tables. Depending on whether each heuristic table depends on locality in the game tree, a memory model might be more suitable than others. This thesis describes an experiment that shows that local tables are generally preferable to global tables for two game heuristics used in chess-like games: killer move and best move history. The experiment is evidence that local tables might be useful for multithreaded applications, particularly ones that involve caching and exhibit locality.