Fast seed-learning algorithms for games

Abstract

Recently, a methodology has been proposed for boosting the computational intelligence of randomized game-playing programs. We propose faster variants of these algorithms, namely rectangular algorithms (fully parallel) and bandit algorithms (faster in a sequential setup). We check the performance on several board games and card games. In addition, in the case of Go, we check the methodology when the opponent is completely distinct to the one used in the training.