Move Evaluation in Go Using Deep Convolutional Neural Networks
Authors: Chris Maddison, Aja Huang, Ilya Sutskever, and David Silver
ICLR 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We train a large 12-layer convolutional neural network by supervised learning from a database of human professional games. The network correctly predicts the expert move in 55% of positions, equalling the accuracy of a 6 dan human player. When the trained convolutional network was used directly to play games of Go, without any search, it beat the traditional-search program Gnu Go in 97% of games, and matched the performance of a state-of-the-art Monte-Carlo tree search that simulates two million positions per move. |
| Researcher Affiliation | Collaboration | Chris J. Maddison University of Toronto cmaddis@cs.toronto.edu Aja Huang1, Ilya Sutskever2, David Silver1 Google Deep Mind1, Google Brain2 {ajahuang,ilyasu,davidsilver}@google.com |
| Pseudocode | No | The paper describes its methods in prose, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access information (e.g., specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described. |
| Open Datasets | No | The dataset used in this work comes from the KGS Go Server. It consists of sequences of board positions st for complete games played between humans of varying rank. We collected 29.4 million board-state next-move pairs (st, at) corresponding to 160,000 games. While the source is named, no specific link, DOI, formal citation with authors/year, or repository for the collected dataset is provided for public access. |
| Dataset Splits | Yes | The dataset was split into a training set of 27.4 million board-state next-move pairs and a test set of 2 million. This split was done before shuffling, so this corresponds to a test set with distinct games. |
| Hardware Specification | Yes | Using a single machine with Intel Xeon CPU E5-2643 v2 @ 3.50GHz and Ge Force GTX Titan Black GPU, we are able to maintain a MCTS search at approximately 47,000 rollouts per second, without dropping CNN evaluations. |
| Software Dependencies | No | The paper mentions methods like 'asynchronous stochastic gradient descent (Dean et al., 2012)' but does not provide specific software names with version numbers (e.g., 'PyTorch 1.9', 'CUDA 11.1') for libraries or frameworks used in the implementation. |
| Experiment Setup | Yes | We used a deep convolutional neural network with 12 weight matrices for each of 12 layers and rectified linear non-linearities. The first hidden layer s filters were of size 5 5 and the remainder were of size 3 3, with a stride of 1. ... For training the network, we used asynchronous stochastic gradient descent (Dean et al., 2012) with 50 replicas each on its own GPU. All parameters were initialized randomly from a uniform[-0.05, 0.05]. Each replica was trained for 25 epochs with a batchsize of 128, a fixed learning rate of 0.128 normalized by batchsize, and no momentum. |