One of the most important and challenging research directions in artificial intelligence is incomplete information games. Mahjong is a typical representative of incomplete information games. Compared with other incomplete information games such as two-player Texas Hold'em and bridge, mahjong is characterized by many players and high uncertainty in the order and strategy of each player, which makes it difficult for the opponent to make reasonable behavioral decisions by accurately assessing the game situation. In this paper, we propose a multi-strategy valuation model based on popular mahjong: by evaluating the winning rate of the situation and the prediction of the opponent's listening tile, we divide the playing strategies into offensive, defensive and general strategies, and adopt different joint playing models for different strategies. The experimental results show that the proposed strategy has a higher level of decision making under a limited number of hands compared with the previous method of single-play model.

为解决当前竞技二打一比赛中存在的赛制冗长、存在作弊隐患的问题,设计一种牌力评估与同等牌力生成系统.对大量人类打牌数据使用二阶聚类算法进行聚类分析,得到手牌牌力分类的标注数据集;构建基于注意力机制的长短期记忆网络(long short term memory network,LSTM)模型,针对手牌序列与手牌牌力进行训练,并通过生成同等牌力手牌借助AI机器人进行对打实验.实验结果表明:生成的不同牌力手牌在胜率上具有显著区别,可正确生成对应牌力手牌,具有可行性.

为利用增强现实(augmented reality,AR)技术实现视力自查互动系统,使用前端框架、canvas 2d渲染引擎和TensorFlow.JS等技术,并设计系统架构和互动测试视力流程,设计实现模型训练页面、互动测试UI部分及对应的动画元素.通过摄像头捕获实时的人物动作,实现系统基于AR的简单互动,系统可以识别用户的表情与手势组合符合哪种预先设定组合,并做出字母方向识别正确与否的判断;通过屏显设备在真实视力测试的场景内,实现AR互动式测试视力的效果,经评估该系统在拥有良好用户体验的情况下实现准确判断.结果表明:通过探究AR技术实现视力自查互动系统,可使用户随时随地自查视力来爱护眼睛,保护视力,并取得了良好的效果.

As a chess species with a long history, research on drafts is mostly limited to the improvement of the chessboard representation method and the static evaluation function. Due to the limitations of the static evaluation function itself, there are problems such as insufficient evaluation accuracy and low judgment accuracy. This paper summarizes and analyzes the existing checkers evaluation algorithm and proposes an evolutionary neural network-based checkers evaluation algorithm. First, the neural network is used to train on different datasets, and the models trained on different datasets are selected as the parent. The model network, through the game between the parent and the offspring, selects the excellent evaluation neural network to keep as the next generation of the parent and iterates continuously and improves the accuracy of the situation evaluation by controlling the step length of the new mutation parameter of the search neural network, thereby get a good position evaluation function. Experiments show that its effect is better than the method based on static situation evaluation.

机器博弈作为人工智能学科中的一个重要领域,是检测人工智能发展水平的一个重要平台.在机器博弈领域中,斗地主游戏具有典型的代表性.本文针对斗地主游戏中的残局策略进行研究,实现了基于Alpha-Beta剪枝算法的残局智能AI.在微信平台小程序"欢乐斗地主"中进行了大量双人明牌斗地主残局测试,均成功获胜,实验表明本文所实现的斗地主残局策略是可行的.

Aiming at the problems of large state space, many game actions and complex game information of Xuezhan Mahjong, a method of semantic segmentation and feature extraction of Xuezhan Mahjong situation is proposed, and a Xuezhan Mahjong decision system based on DenseNet network and human empirical knowledge is designed and implemented according to the extracted game situation. The experimental results show that the model produces good training results in the original data and has high accuracy on the test set. The proposed Discard model outperforms CNN, ResNet, and XGBoost Discard models, and the decisions made by the system are human-like in playing with real players. It provides a research basis for the subsequent Xuezhan Mahjong Discard model that can surpass human.

The imperfect information game has too much hidden information in the game process, which leads to the huge information set in the game process. It is difficult to solve the problem through the mature game search algorithm. This paper takes the game guessing as the carrier, imitates the game behavior of human beings in the game remnant state, and gives a method to transform the imperfect information game into the perfect information game. Through this method, the other players' hands in the game are guessed, and the imperfect information is transformed into perfect information, so as to solve the problem that the data dimension brought by the hidden information of the game is too large Questions. As for the game of DouDizhu, in the end of game, there are 133 kinds of ending states. In this paper, the game information generated in the process of DouDizhu is studied, the appropriate representation method of game information is given, and a reasonable 3D time series convolution model is designed to extract features, and then a classification model with high accuracy of guessing cards is obtained. The model has a high accuracy of 92.1% for the final game. The perfect game search algorithm is used to solve the game's hand out strategy to improve the game AI's playing strength.

传统的人体骨骼动画制作方法是参考真实人体骨骼运动过程中各关节点的坐标位置变化轨迹,确定关键帧与插值帧模型,再进行必要的编辑处理,这需要大量的专业领域知识以及复杂的交互规则,人力物力成本高.为解决上述问题,提出了一种在给定任意两个人体骨骼动作数据作为首尾帧的条件下,自动合成完整三维人体骨骼动画序列的方法.方法由基于卷积自编码网络的骨骼动画特征提取模型和双线性插值与卷积相结合的插值帧自动合成模型两部分组成.随机从Mocap数据库中抽取两帧人体动作数据作为模型的输入,可以自动生成三维人体骨骼动画.文中方法与传统插值帧生成方法相比,具有较好的动作趋势转折信息的预测和还原能力,提升了动画系统的交互效率以及智能水平.

This paper studies the current situation of computer game platforms and focuses on the rational use and expansion of existing game technology to build a complete general-purpose computer game platform that can load different game engines. The front-end and back-end decoupling architecture of the game platform was designed. The front-end calls the Native API at the back-end through WebSocket and uses JSON data for interaction. Realize the role of the game platform in the game rule determination and input–output interaction interface, and provide higher execution efficiency, more convenient operation mode and necessary rule evaluation for game participants. Let computer games be more fair, just and efficient. Experiments show that the system is feasible and effective, and it has practical significance for the future study of computer game.

全国竞技二打一扑克锦标赛是国家体育总局棋牌运动管理中心推出的新兴项目,比赛采用复式赛赛制,然而这种赛制存在无法杜绝线上比赛作弊的问题,为此提出用同等牌力赛制来代替复式赛制进行比赛的设想.本文主要针对如何判定和选取具有同等牌力的初始手牌进行研究,提出用等级难度评估指标来计算初始手牌的难度,初始手牌难度相近的牌认定为同等牌力,并给出了同等牌力的选取方法;为保证一副牌所得等级难度评估指标的可靠性,提出用多个不同水平的"二打一AI程序"来代替真人打牌产生牌谱的方法.实验证明了同等牌力生成方法的有效性,本文提出的同等牌力方法应用于比赛是可行的.

As an important field of artificial intelligence, computer game is an important platform for testing the development level of artificial intelligence. In the field of incomplete information of computer games, Dou dizhu game is typical. Bidding is a key link in the game, in which the strategy of bidding plays an important role in the outcome of the game. In this paper, the problem of bidding in Dou dizhu game is studied, and the definition of points representing the quality of hand card is given. Firstly, according to the hand card information, the split priority is set according to different card types, so as to obtain different card splitting schemes, and the hand score of each card splitting scheme is calculated. The scheme with the lowest score is the optimal card splitting scheme. Then, according to this score, the corresponding bidding is given. The experiment is compared with a large number of real-life data, and the results show that the matching degree is 85%, which shows that the scheme proposed in this paper is feasible.

针对全国竞技二打一扑克锦标赛中如何保证比赛的公平、公正性,降低冗长的比赛赛程问题,提出通过给出具有较高区分度的初始手牌用于比赛发牌阶段以减少比赛轮次的解决方案,并提出计算初始手牌区分度的算法.首先对大量真实用户打牌行为数据进行分析及处理,计算每一个初始手牌对应的得分概率分布;再计算由若干副该初始手牌组成的初始手牌组的区分度,并进行非降序排序;选取区分度值高者对应的初始手牌作为测试用例.为测试选取用例的有效性,创造性提出采用多个具有不同区分度的"智能二打一竞技者"进行模拟比赛.实验结果表明测试用例的选取方法有效.

The Dou Dizhu game is a popular game in China, which has been accepted as a competition in China Computer Game Championship in recent years. The key step of the game is to bid the card, which directly determines the result of the game. In this paper, the related techniques of deep learning have been deeply studied, and the Dou DiZhu game bidding strategy based on deep learning is put forward. Basing on the real game data of a famous online game company in China, this paper gives an innovative method of data calibration and a card representation model and selects the CNN model to train the bidding model. Finally, the bidding system will give a score according to the given initial hand card, and the program is used as AI program to participate in the Chinese University Computer AI Game Championship in 2020. The results show that the bidding strategy proposed in this paper is feasible.

基于深度学习模型的有监督训练依赖于大量高质量标定数据,针对非完全信息博弈中二打一智力游戏问题,根据不同阶段回合局面数据的特点,提出了通过Alpha-Beta完全搜索获得共包含400万带标定二打一智力游戏局面样本的数据集,根据得到的标定样本训练CNN模型,使其能够对二打一智力游戏残局进行局面评估,为进一步将牌类游戏向棋类游戏的转化提供了保障,也为其他非完全信息博弈训练数据的标定提供了有价值的借鉴.

Tibetan Jiuqi is a two-person chess type of Chinese ethnic minorities. In 2019, it was listed as one of the chess types of China Computer Game Championship. There are relatively few researches on Jiuqi's game strategy at home and abroad. The layout of the pieces of Jiuqi is generally divided into three stages: the opening, the middle game, and the endgame. This article is mainly based on the rules of the game to study the shape of the game in the opening and middle games. In the opening stage, the opening library is used to assist the game. In the middle stage, a strategy of forming a chess shape is proposed. This strategy is based on a single Dalian and further derives a more aggressive and flexible chess shape on this basis. The experimental results show that the strategy proposed in this article is effective.

本文探讨了基于蒙特卡洛方法的强化学习博弈程序的原理,基于该原理结合BP算法设计了一个进行自学习的苏拉卡尔塔博弈程序.实验证明,该方法能让智能体不断的学习提高棋力,避免了繁琐的手工构建静态评估函数过程.

为了提高三维人体骨骼模型的建模效率并简化交互规则,提出了一种基于深度学习的手绘人体动作草图到三维骨骼模型的重建方法.首先将三维骨骼模型渲染为二维图像来建立维度映射关系,进而使用图像分类方法识别手绘草图动作并根据维度映射实现三维骨骼模型重建.在实验中使用基于深度卷积神经网络对图像分类模型进行构建,使用浅层卷积网络作为训练单元,并使用逐级分类与分块训练策略加速网络收敛速度来提高训练效率.最后实验结果验证了该方法的可行性与有效性.

Doudizhu poker is a very popular and interesting national poker game in China, and now it has become a national competition in China. As this game is a typical example of incomplete information game problem, it has received more and more attention from artificial intelligence experts. This paper proposes a multirole modeling-based card-playing framework. This framework includes three parts: role modeling, cards carrying, and decision-making strategies. Role modeling learns different roles and behaviors by using a convolutional neural network. Cards carrying can calculate reasonable rules especially for “triplet” by using an evaluation algorithm. Decision making is for implementing different card strategies for different player roles. Experimental results showed that this card-playing framework makes playing decisions like human beings, and it can to some extent learn, collaborate, and reason when facing an incomplete information game problem. This framework won the runner-up in the 2018 China Computer Game Competition.

This paper studies the current situation of computer game platforms, summarizes the advantages and disadvantages of existing platforms, and proposes to build an interactive computer game platform with a personalized interface based on the Electron front-end framework, using JavaScript, HTML and CSS technology stacks. The system integrates the characteristics of different chess games, designs a reasonable data structure representing the chessboard and chess pieces, and establishes a game model for specific chess types. The interface design of the battle platform follows certain specifications. In addition to the regular menu interface and game battle interface, the corresponding functions are also designed and realized, such as timer, replaying disk, and penitence chess. The chess motion animation display is realized and the chess game is controlled according to the chess rules. By using the computer game platform, the game operation and rule judgment can be realized automatically. Experiments show that the system is feasible and effective, and it has practical significance for the future research of computer games.

Situation assessment and search are two key problems in computer game research. In general, as the game progresses, the difficulty of evaluating the situation of the game is significantly reduced, and the accuracy of the evaluation is significantly increased. Based on the famous chess game, this article proposes and implements a new scheme that combines the Monte Carlo tree search algorithm, the Alpha-Beta algorithm and the model based on the deep convolution neural network (CNN) to solve the computer game problem. This article first proposes a deep convolutional neural network model based on dots and boxes, including deep value network and deep strategy network, focusing on situation assessment and strategy recommendation, respectively. Then, using the Monte Carlo Tree Search (MCTS) algorithm as a framework, deep value network integrated MCTS algorithm and deep strategy network integrated MCTS algorithm are proposed. In both integrated models, Alpha-Beta complete search is used to truncate the Monte Carlo simulation process and improve simulation efficiency. Through competition with human players, the results show that the two integrated algorithm game systems have reached much higher intelligence level than ordinary humans in solving the problem of dots and boxes.