Training Deep Convolutional Neural Networks to Play Go

Christopher Clark, Amos Storkey
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:1766-1774, 2015.

Abstract

Mastering the game of Go has remained a long-standing challenge to the field of AI. Modern computer Go programs rely on processing millions of possible future positions to play well, but intuitively a stronger and more ’humanlike’ way to play the game would be to rely on pattern recognition rather than brute force computation. Following this sentiment, we train deep convolutional neural networks to play Go by training them to predict the moves made by expert Go players. To solve this problem we introduce a number of novel techniques, including a method of tying weights in the network to ’hard code’ symmetries that are expected to exist in the target function, and demonstrate in an ablation study they considerably improve performance. Our final networks are able to achieve move prediction accuracies of 41.1% and 44.4% on two different Go datasets, surpassing previous state of the art on this task by significant margins. Additionally, while previous move prediction systems have not yielded strong Go playing programs, we show that the networks trained in this work acquired high levels of skill. Our convolutional neural networks can consistently defeat the well known Go program GNU Go and win some games against state of the art Go playing program Fuego while using a fraction of the play time.

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-clark15, title = {Training Deep Convolutional Neural Networks to Play Go}, author = {Clark, Christopher and Storkey, Amos}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {1766--1774}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/clark15.pdf}, url = { http://proceedings.mlr.press/v37/clark15.html }, abstract = {Mastering the game of Go has remained a long-standing challenge to the field of AI. Modern computer Go programs rely on processing millions of possible future positions to play well, but intuitively a stronger and more ’humanlike’ way to play the game would be to rely on pattern recognition rather than brute force computation. Following this sentiment, we train deep convolutional neural networks to play Go by training them to predict the moves made by expert Go players. To solve this problem we introduce a number of novel techniques, including a method of tying weights in the network to ’hard code’ symmetries that are expected to exist in the target function, and demonstrate in an ablation study they considerably improve performance. Our final networks are able to achieve move prediction accuracies of 41.1% and 44.4% on two different Go datasets, surpassing previous state of the art on this task by significant margins. Additionally, while previous move prediction systems have not yielded strong Go playing programs, we show that the networks trained in this work acquired high levels of skill. Our convolutional neural networks can consistently defeat the well known Go program GNU Go and win some games against state of the art Go playing program Fuego while using a fraction of the play time.} }
Endnote
%0 Conference Paper %T Training Deep Convolutional Neural Networks to Play Go %A Christopher Clark %A Amos Storkey %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-clark15 %I PMLR %P 1766--1774 %U http://proceedings.mlr.press/v37/clark15.html %V 37 %X Mastering the game of Go has remained a long-standing challenge to the field of AI. Modern computer Go programs rely on processing millions of possible future positions to play well, but intuitively a stronger and more ’humanlike’ way to play the game would be to rely on pattern recognition rather than brute force computation. Following this sentiment, we train deep convolutional neural networks to play Go by training them to predict the moves made by expert Go players. To solve this problem we introduce a number of novel techniques, including a method of tying weights in the network to ’hard code’ symmetries that are expected to exist in the target function, and demonstrate in an ablation study they considerably improve performance. Our final networks are able to achieve move prediction accuracies of 41.1% and 44.4% on two different Go datasets, surpassing previous state of the art on this task by significant margins. Additionally, while previous move prediction systems have not yielded strong Go playing programs, we show that the networks trained in this work acquired high levels of skill. Our convolutional neural networks can consistently defeat the well known Go program GNU Go and win some games against state of the art Go playing program Fuego while using a fraction of the play time.
RIS
TY - CPAPER TI - Training Deep Convolutional Neural Networks to Play Go AU - Christopher Clark AU - Amos Storkey BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-clark15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 1766 EP - 1774 L1 - http://proceedings.mlr.press/v37/clark15.pdf UR - http://proceedings.mlr.press/v37/clark15.html AB - Mastering the game of Go has remained a long-standing challenge to the field of AI. Modern computer Go programs rely on processing millions of possible future positions to play well, but intuitively a stronger and more ’humanlike’ way to play the game would be to rely on pattern recognition rather than brute force computation. Following this sentiment, we train deep convolutional neural networks to play Go by training them to predict the moves made by expert Go players. To solve this problem we introduce a number of novel techniques, including a method of tying weights in the network to ’hard code’ symmetries that are expected to exist in the target function, and demonstrate in an ablation study they considerably improve performance. Our final networks are able to achieve move prediction accuracies of 41.1% and 44.4% on two different Go datasets, surpassing previous state of the art on this task by significant margins. Additionally, while previous move prediction systems have not yielded strong Go playing programs, we show that the networks trained in this work acquired high levels of skill. Our convolutional neural networks can consistently defeat the well known Go program GNU Go and win some games against state of the art Go playing program Fuego while using a fraction of the play time. ER -
APA
Clark, C. & Storkey, A.. (2015). Training Deep Convolutional Neural Networks to Play Go. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:1766-1774 Available from http://proceedings.mlr.press/v37/clark15.html .

Related Material