Transformer models have demonstrated impressive capabilities when trained atscale, excelling at difficult cognitive tasks requiring complex reasoning andrational decision-making. In this paper, we explore the application oftransformer models to chess, focusing on the critical role of the positionencoding within the attention mechanism. We show that in chess, transformersendowed with a sufficiently versatile position encoding can match existingchess-playing models at a fraction of the computational cost. Our architecturesignificantly outperforms AlphaZero at 8x fewer FLOPS and matches priorgrandmaster-level transformer-based agents at 30x fewer FLOPS. <<<

The AlphaZero algorithm has been successfully applied in a range of discretedomains, most notably board games. It utilizes a neural network, that learns avalue and policy function to guide the exploration in a Monte-Carlo TreeSearch. Although many search improvements have been proposed for Monte-CarloTree Search in the past, most of them refer to an older variant of the UpperConfidence bounds for Trees algorithm that does not use a policy for planning.We introduce a new, improved search algorithm for AlphaZero which generalizesthe search tree to a directed acyclic graph. This enables information flowacross different subtrees and greatly reduces memory consumption. Along withMonte-Carlo Graph Search, we propose a number of further extensions, such asthe inclusion of Epsilon-greedy exploration, a revised terminal solver and theintegration of domain knowledge as constraints. In our evaluations, we use theCrazyAra engine on chess and crazyhouse as examples to show that these changesbring significant improvements to AlphaZero. <<<

This paper introduces ZeusAI, an artificial intelligence system developed toplay the board game 7 Wonders Duel. Inspired by the AlphaZero reinforcementlearning algorithm, ZeusAI relies on a combination of Monte Carlo Tree Searchand a Transformer Neural Network to learn the game without human supervision.ZeusAI competes at the level of top human players, develops both known andnovel strategies, and allows us to test rule variants to improve the game'sbalance. This work demonstrates how AI can help in understanding and enhancingboard games. <<<