PDF] Monte-Carlo Graph Search for AlphaZero

A new, improved search algorithm for AlphaZero is introduced which generalizes the search tree to a directed acyclic graph, which enables information flow across different subtrees and greatly reduces memory consumption. The AlphaZero algorithm has been successfully applied in a range of discrete domains, most notably board games. It utilizes a neural network, that learns a value and policy function to guide the exploration in a Monte-Carlo Tree Search. Although many search improvements have been proposed for Monte-Carlo Tree Search in the past, most of them refer to an older variant of the Upper Confidence bounds for Trees algorithm that does not use a policy for planning. We introduce a new, improved search algorithm for AlphaZero which generalizes the search tree to a directed acyclic graph. This enables information flow across different subtrees and greatly reduces memory consumption. Along with Monte-Carlo Graph Search, we propose a number of further extensions, such as the inclusion of Epsilon-greedy exploration, a revised terminal solver and the integration of domain knowledge as constraints. In our evaluations, we use the CrazyAra engine on chess and crazyhouse as examples to show that these changes bring significant improvements to AlphaZero.

PDF) Targeted Search Control in AlphaZero for Effective Policy Improvement

Mastering the game of Go without human knowledge

Global optimization of quantum dynamics with AlphaZero deep exploration

Acquisition of chess knowledge in AlphaZero

PDF] Monte-Carlo Graph Search for AlphaZero

Monte Carlo Tree Search: a review of recent modifications and applications

AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Shogi and Go through Self-Play

Monte-Carlo Graph Search for AlphaZero

Monte Carlo tree search - Wikipedia