Training Deep Convolutional Neural Networks to Play Go

Authors: Christopher Clark, Amos Storkey

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 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.
Researcher Affiliation Collaboration Christopher Clark CHRISC@ALLENAI.ORG Allen Institute for Artificial Intelligence , 2157 N Northlake Way Suite 110, Seattle, WA 98103, USA Amos Storkey A.STORKEY@ED.AC.UK School of Informatics, University of Edinburgh, 10 Crichton Street, Edinburgh, EH9 1DG, United Kingdom
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes We use two datasets to evaluate our approach. The first is the Games of Go on Disk2 (Go Go D) dataset consisting of 81,000 professional Go games. [...] The second dataset consists of 86,000 Go games played by high ranked players on the KGS Go server 34. [...] 2http://gogodonline.co.uk/ 3https://www.gokgs.com/ 4http://u-go.net/gamerecords/
Dataset Splits Yes Both datasets were partitioned into test, train, and validation sets each containing position-move pairs that are from disjoint games of Go. We use 8% (1.3 million) for testing, 4% (620 thousand) for validation, and the rest (14.7 million) for training.
Hardware Specification Yes The network was trained for seven epochs at a learning rate of 0.05, two epochs at 0.01, and one epoch at 0.005 with a batch size of 128 which took roughly four days on a single Nvidia GTX 780 GPU.
Software Dependencies No The paper mentions 'GNU Go 3.8' and 'Feugo 1.1' as opponents, but does not provide specific version numbers for the software libraries or frameworks used to implement their own deep convolutional neural networks.
Experiment Setup Yes The network was trained for seven epochs at a learning rate of 0.05, two epochs at 0.01, and one epoch at 0.005 with a batch size of 128 which took roughly four days on a single Nvidia GTX 780 GPU. [...] Both convolutional and fully connected layers had their biases initialized to zero and weights drawn from a normal distribution with mean 0 and standard deviation 0.01.