SDM-Go: An Agent for Go with an Improved Search Process Based on Monte-Carlo Tree Search and Sparse Distributed Memory

This paper describes the player agent for Go named SDM-Go. It uses sparse distributed memory (SDM) as an additional resource to improve the classical Monte-Carlo (MC) simulation based search used by many current top player agents. In this process, the search tree used to select the best move is constructed according to the values of the game boards. The SDM-Go is built upon the Fuego framework. The contributions here are: the use of a SDM to reduce the number of MC simulations and to increase the accuracy of the calculus of the values corresponding to the game boards (nodes) that keep a certain level of similarity with some node of the SDM, the implementation of a bit based representation for the game boards such as not to compromise the performance of the system when checking the similarities among them and, to extend the use of the results of the simulations to update the values of the nodes of the SDM. The use of the SDM represents an approach independent of domain, what reduces the supervised character of the traditional MC simulations. The results obtained in tournaments against the well known open-source agent Fuego confirm the benefits provided by this approach.