【作者】 布伟光；

【导师】 何中市；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2008， 硕士

【摘要】 近年来,游戏的图形质量已发展到近乎极致的地步,人工智能(AI)已经成为决定一款游戏以及游戏开发工作室命运的重要因素。下一代的3D游戏不仅会有优秀的视觉效果,更会像人一样狡猾和聪明。由于国内还没有展开全面研究和应用,而且本文研究并实现的AI技术在一定程度上提高了游戏智能,因此该课题具有一定的学术和应用价值。本文目标是构建一个基本图形渲染引擎,以这个渲染引擎为平台,对若干基于规则和机器学习的AI技术进行了深入研究和实现,并应用若干机器学习技术实现了游戏中一些常见问题的求解。首先构建图形渲染引擎,功能有:完整流水线,物体剔除,背面消除,欧拉相机,光照模型,固定、恒定和Gouraud着色方法,3D裁剪,深度排序。然后研究实现了若干基于规则的AI技术。基于规则的AI技术包括:确定性运动算法、随机运动算法、跟踪闪避算法、群聚算法、模式运动技术、行为建模的有穷自动机技术和A*算法等。最后采用机器学习中的遗传算法和人工神经网络,实现了游戏中以下问题的求解:1.寻路问题。一条染色体代表一条路径。实验表明遗传算法对结果的不可预知性可以有效地提高寻路的智能;2.飞行物体的着陆问题。染色体由飞行物体的运行方式组成。实验表明遗传算法对结果的不可预知性使得降落更加智能化,不需要人工的干预;3.障碍物绕行问题。使用遗传算法改进神经网络的权值;利用以智能体中心为出发点的5条射线模拟传感器感知环境。经过768代的进化,遗传算法种群最优适应度和平均适应度都有了明显提高,绕行成功率从12.5%上升到85%;4.鼠标轨迹识别问题。神经网络权值由反向传播算法学习更新;采用1200个样例训练,误差阈值为34.0037,另外1200个样例测试,将神经网络与SVM做了实验对比,得出结论:神经网络和SVM的正确分类样本数分别为1125和1185,错误分类样本数分别为75和15,正确率分别为93.75%和98.75%。下一步工作,希望将这些AI技术有机地结合起来,整合为一个AI引擎,应用在实际的游戏项目中。更多还原

【Abstract】 Recent years, because of perfect development of graphic hardware and rendering quality in video games, artificial intelligence has become a more and more important factor for success of game developing studios. Next generation of 3D video games not only has perfect optical effect, but also advanced AI.Due to lack of game AI papers and applications, these techniques in this paper, which can improve AI level in games indeed, achieve a measure of academic and applied values.The target of paper is that building a basic graphic rendering engine first, then based on the engine, several Rules-Based AI techniques are researched and implemented, finally, some usual problems in games are resolved by some machine learning techniques.Firstly, the graphic rendering engine has an whole pipeline, the functions of the engine as follows: Object culling, Back-face removing, Euler camera model, Lighting model, Constant shading, Flat shading, Gouraud shading, 3D clipping and Depth sorting.Secondly, some Rules-Based AIs are researched. These Rules-Based AIs include: Deterministic movement algorithm, random movement algorithm, chasing (evading) algorithm,Flocking algorithm, Pattern movement technique, FSM technique and A* algorithm.Finally, These 4 problems are resolved by using Genetic Algorithm(GA) and Artificial Neural Network(ANN):1. Pathfinding.The experiment shows that the unpredictable feature of GA can improve intelligence of Pathfinding.2. spaceship landing. Chromosome consists of movement pattern of spaceship. The result shows that the unpredictable feature of GA can improve intelligence of spaceship landing, no artificial controlling needed.3. obstacle avoidance. ANN’s weights are updated by GA. Sensors are simulated by 5 line segments that radiate outward from the agent body, the agent can sense the game environment by the 5 sensors. After iterating for 768 times, average fitness and best fitness of the population have been improved quickly, the ratio of avoiding successfully has been improved from 12.5% to 85%.4. Resolved mouse track recognition. ANN’s weights are updated by Back Propagation algorithm. The size of training data set is 1200, error threshold is 37.0037, the size of testing data set is 1200, according to contrastive analysis of ANN and SVM, the result is that the right recognition number of testing samples for ANN is 1125, for SVM is 1185, the wrong recognition number of testing samples for ANN is 75, for SVM is 15, the precision for ANN is 93.75%, for SVM is 98.75%.Next step,AI techniques in this paper will be integrated into an AI engine,and this AI engine is expected to actual game projects.更多还原