Adversarial Search Algorithms Performance in the Yote Game

Abstract

The performance of several adversarial search algorithms is very dependent on the move generation and ordering procedure, as well as on the transposition table characteristics. In this work we compare Alpha-beta, MTD-f and NegaScout performance in the presence and absence of the following features: transposition table, Hash Move heuristic and move ordering using the evaluation function. We found that NegaScout did not yield major improvements over Alpha-Beta, which is likely due to the move ordering performed using a weak evaluation function. In most cases, MTD-f examines fewer nodes, and completes fixed depth searches in less time than Alpha-beta and NegaScout. This is consistent with the literature when using coarse grained evaluation functions. Finally, we found that enabling move ordering leads to fewer nodes examined than enabling the Hash Move heuristic, but, as sorting is time consuming, search is slower. Best results overall are found using MTD-f with all enhancements enabled. These results, in a previously unexplored game, are consistent with existing literature. This enables us to make predictions of algorithm and heuristics’ performance in novel games, based on an analysis of the games’ features, without having to implement all techniques.