A Self–Trained Engine for Atomic Chess

Abstract

This thesis concerns the development of a chess engine to play a variant of chess called atomic chess, utilizing a neural network. The neural network is modeled after DeepMind’s AlphaZero, which is a model that learned standard chess from only the rules, and no real-world games (hence “zero”). We demonstrate an adapted model that improves its playing strength in atomic chess, given enough training time. This is done using a deep convolutional neural network, trained with data generated by a modified Monte Carlo tree search in a process called self-play. These two components feed each other data in a cycle: the neural network guides the tree search, and the results of the tree search are then used to train the network, repeatedly. The result is a trained network that is shown to have improved from the untrained model, which corresponds to an unguided or randomly guided Monte Carlo tree search. An extensive background aimed at computer engineering students is also included, explaining the terms used in the thesis.