5 February 2018 in IEEE Transactions on Games
Endgame databases can be difficult to use in tree search when database sizes remain large even after compression. Given the same endgame, we discover that the compression ratios vary significantly when using different encoding schemes. The intuition is that when a set of positions mapped into a continuous chunk of segments have similar values, block-based compression libraries such as zlib can yield a better compression ratio than cases where segments contain diversified values. However, finding the optimal encoding scheme by exhaustive enumeration is time-infeasible for endgame databases with a large number of pieces. We propose a novel approach using deep learning to obtain an encoding scheme so that the compression ratio is suitable for practical purposes. Our approach can be applied to chess-like games.
16 January 2019 in IEEE Transactions on Games
This paper describes the application of a modified comparison training for automatic feature weight tuning. The final objective is to improve the evaluation functions used in Chinese chess programs. First, we apply n-tuple networks to extract features. N-tuple networks require very little expert knowledge through its large numbers of features, while simultaneously allowing easy access. Second, we propose a modified comparison training into which tapered eval is incorporated. Experiments show that with the same features and the same Chinese chess program, the automatically tuned feature weights achieved a win rate of 86.58% against the hand-tuned ones. The above trained version was then improved by adding additional features, most importantly n-tuple features. This improved version achieved a win rate of 81.65% against the trained version without additional features.
29 July 2019 in IEEE Transactions on Games
Deep learning has become a hot topic in recent years. There are many teaching frameworks that ease the education process for deep learning. However, most current teaching examples either require a lot of training time or do not have interaction with users. Usually, the testing accuracy is the only evaluation criterion. However, it does not mean too much for a novice. We provide a framework which can teach the deep learning concepts with limited time or computational resource constraints. It offers students quick feedback, that is, the trained model can be easily tested on this framework. Students can increase training data, design a deep learning model or modify input planes to learn more advanced topics. The framework provides a Graphical User Interface, which connects to the trained model directly. It is a well-designed deep learning competition framework. This framework has been applied in the “Artificial Intelligence” undergraduate course for three semesters. The student feedback was quite positive. Furthermore, based on this framework, Deep Learning Othello competitions were held in TAAI and TCGA international computer game tournaments in 2017. We hope that this framework can encourage more people to be engaged in the research of deep learning and board games.
21 February 2023 in IEEE Access
The AlphaZero algorithm achieved superhuman levels of play in chess, shogi, and Go, by learning without domain-specific knowledge except for game rules. This paper targets stochastic games and investigates whether AlphaZero can learn theoretical values and optimal play. Since the theoretical values of stochastic games are expected win rates, not a simple win, loss, or draw, it is worth investigating the ability of AlphaZero to approximate expected win rates of positions. This paper also thoroughly studies how AlphaZero is influenced by hyper-parameters and some implementation details. The analyses are mainly based on AlphaZero learning with lookup tables. Deep neural networks (DNNs) like the ones in the original AlphaZero are also experimented and compared. The tested stochastic games include reduced and strongly-solved variants of Chinese dark chess and EinStein würfelt nicht!. The experiments show that AlphaZero can learn policies that play almost optimally against the optimal player and can learn values accurately. In more detail, such good results are achieved by different hyper-parameter settings in a wide range, though it is observed that games on larger scales tend to have a little narrower range of proper hyper-parameters. In addition, the results of learning with DNNs are similar to lookup tables.