0467760 - ÚI 2017 RIV CH eng C - Konferenční příspěvek (zahraniční konf.)
Moudřík, Josef - Neruda, Roman
Determining Player Skill in the Game of Go with Deep Neural Networks.
Theory and Practice of Natural Computing. Cham: Springer, 2016 - (Martín-Vide, C.; Mizuki, T.; Vega-Rodríguez, M.), s. 188-195. Lecture Notes in Computer Science, 10071. ISBN 978-3-319-49000-7. ISSN 0302-9743.
[TPNC 2016. International Conference on the Theory and Practice of Natural Computing /5./. Sendai (JP), 12.12.2016-13.12.2016]
Grant CEP: GA ČR GA15-19877S
Institucionální podpora: RVO:67985807
Klíčová slova: computer Go * machine learning * board games * skill assessment * deep neural networks
Kód oboru RIV: IN - Informatika

The game of Go has recently been an exuberant topic for AI research, mainly due to advances in Go playing software. Here, we present an application of deep neural networks aiming to improve the experience of humans playing the game of Go online. We have trained a deep convolutional network on 188,700 Go game records to classify players into three categories based on their skill. The method has a very good accuracy of 71.5 % when classifying the skill from a single position, and 77.9 % when aggregating predictions from one game. The performance and low amount of information needed allow for a much faster convergence to true rank on online Go servers, improving user experience for new-coming players. The method will be experimentally deployed on the Online Go Server (OGS).
Trvalý link: http://hdl.handle.net/11104/0265797