【作者】 赵斌；

【导师】 李一民；

【作者基本信息】 昆明理工大学 ， 模式识别与智能系统， 2004， 硕士

【摘要】 多主体系统(MAS)是当前人工智能研究的主攻方向。就是在一个实时动态变化的和不可预测的环境中，有自主能力的自主体，作为团队的一部分，自动地进行有效的动作，完成整个队伍的目标。机器人足球就是这样的一个多主体系统，它具有MAS的所有特性。同时，机器人足球具有观赏性和刺激性，可以很好的展示多主体系统的研究成果。鉴于机器人足球的这些特点，国际人工智能界选择了机器人足球作为多主体系统研究的一个标准问题，在本文中，我们使用仿真机器人足球作为我们研究的载体。 首先，我们设计出了良好的仿真机器人(Agent)程序结构。在消化前人成果的同时，对UvA Trilearn的同步机制作了修订，对它的调试系统作了大量的修改，有效的提高了调试和开发效率。在Agent世界模型中，运用了Kalman滤波对噪声进行了过滤，极大的提高了世界模型精度，给高质量的底层动作打下扎实的基础。在底层动作的设计中，对分析法(几何法)和基于人工神经网络经验法作了深入分析，给出了典型算法。在系统决策中，比较了传统的个人能力层与阵型层实现方法后，引入了所有队员可以同时学习的加强学习方法对动作寻优。虽然加强学习在Soccer Server仿真环境下应用有储多限制，但这是对实现Agent智能行为的有益探索。对传统的阵型层，提出了阵型、战术库，并可以静态动态的配置它们的方法。在对手建模中，创造性的提出了用非线性的人工神经网络来对对手的阵型与战术做辨识，这将极大的提高Agent的智能与球队的比赛成绩。最后为了避免重复的繁重的调试分析劳动，提出了用Matlab来对Soccer Server作仿真，利用Matlab的各种工具箱和统计函数来加速对新思想和新算法的研究应用。更多还原

【Abstract】 Multi-Agent Systems (MAS) is the main subfield of Artificial Intelligence. In complex, real-time and unpredictable systems, it requires agents to act effectively both autonomously and as part of a team, to cooperate with the teammates and defense again the opponents, to achieve the team long-term goal. The robotic soccer is such a multi-agent system, and it is a domain that fits all the above characteristics. At the same time, the robotic soccer is enjoyable and exciting. It can display research production of multi-agent system and robotics. Consider all the characteristics of the robotic soccer, the International Joint Committee of Artificial Intelligence choose robotic soccer as a standard problem of MAS. In this thesis, we use the simulation robotic soccer as test bed.First, we design an effective architecture for agent programming. Digesting former scholar’s work, we correct the synchronization’s bug in UvA Trilearn, and greatly improve it’s debug system, which can effectively help to develop agent program. In agent’s world model, we use Kalman filter to deal with noise, which greatiy improve the precision and give a reliable base for developing high-quality personal skills of agent. To personal skill of agent, we analyze the analysis way(geometry way) and experiential way based on neural networks, and give out typical algorithm. To agent’ system-decision, comparing classical way based two layers (personal ability and formation), we use reinforcement learning that can let all teammates learn synchronously to seek optimal actions. Although reinforcement learning for simulated soccer has many restriction and may not bring the predictive success, it is the direction of agent’s intelligence programming. To team strategy (formation layer), we propose the method to use the formation and tactics set that is dynamic and can be configured flexibly. In opponent modeling, we propose a creative method to use non-linear neural networks to recognize opponent’s strategy and formation, which greatly improve agent’s intelligence and match scores. Lastly, to avoid heavy and boring low-level analysis and debug work, we propose to use Matlab to simulate soccer server, which have advantage of use it’s many tool boxes and many math functions such as statistical functions.更多还原