子計畫四：雙層疊工作層級運算系統應用於外圍開局五子棋及相關程式之研發

The development of Two-Tier JLCS has become increasingly stable, and the AlphaZero's method has shown its great potential, hence, during the implementation of this subproject, we are committed to improving AlphaZero's method and applying AlphaZero's method to Outer-Open Gomoku and other computer games. In this subproject, we have submitted two papers last year that are accepted by CG2018 (The 10th International Conference on Computers and Games). The first one focuses on connection-type games (Tic-Tac-Tao&Outer Opening Gomoku) for experiments. We aim at the characteristics of connection-type games, to propose some new ideas (limiting the search space, giving the depth rewards) to reduce the breadth and depth and hence improve the performance of MCTS methods. We are having a good achievement and are going to do more tests as well as tunings. The report also includes the design issue of the endgame tablebase of the other paper that will be presented in CG2018. In this final report, we further describe the technics used in our Chinese Checkers program. In addition, we have published and presented our improving ideas in many international and domestic journals and conferences. Chinese Checkers is a very popular game but lacks research papers on promoting its strength of computer programs. In past, the Chinese Checkers programs based on Monte-Carlo Tree Search can already show a certain degree of strength, but there is still room for improvement. In order to play well, Chinese Checkers programs need to enable the pieces to move forward quickly, they must also retreat at the appropriate time to achieve a balance between offense and defense. This research is based on Monte Carlo tree search to explore the improved schemes of two-person Chinese Checkers, and adds the methods of self-reinforcement learning as well as the AlphaZero approach to train the neural networks. In order to obtain training results under limited hardware resources and time conditions, this study considers some artificial knowledge to the characteristics of Chinese Checkers, and generates chess records in advance as training materials for self-reinforcement learning. Finally, the idea of moves modification is used to make training more accurate. Experiments have proved that the schemes proposed in this subproject can train a Chinese Checkers program with a considerable strength in a short time. In addition, this subproject is also dedicated to the development of other computer game programs. In the past two years, we participated in five international/domestic computer game competitions. In these five competitions, our research group of this subproject has won 22 gold, 18 silver and 10 bronze medals. We also won the second place in the overall team championship in the TAAI 2017 computer game competition. In terms of paper publication, during the implementation of the subproject, we have published 4 SCIE journal papers, 3 EI international conference papers, and 4 domestic conference papers. At present, two EI international conference papers and one domestic conference paper have been accepted, and another SCIE journal paper is under review. Based on the above points, the subproject has achieved the expected progress and goals.