Pruning Stochastic Game Trees Using Neural Networks for Reduced Action Space Approximation

Abstract

Monte Carlo Tree Search has proved to be very efficient in the broad domain of Game AI, though it suffers from high dimensionality in cases of large branching factors. Several pruning techniques have been proposed to tackle this problem, most of which require explicit domain knowledge. In this study, an approach using neural networks to determine the number of actions to be pruned, depending on the iterations run and the total number of possible actions, is proposed. Multi-armed bandit simulations with the UCB1 formula are employed to generate suitable datasets for the networks’ training and a specifically designed process is followed to select the best combination of the number of iterations and actions for pruning. Two pruning Monte Carlo Tree Search variants are investigated, based on different actions’ expected rewards’ distributions, and they are evaluated in the collectible card game Hearthstone. The proposed technique improves the performance of the Monte Carlo Tree Search algorithm in different setups of computational limitations regarding the available number of tree search iterations and is significantly boosted when combined with supervised learning trained-state value predicting models.