【作者】 黄颖；

【导师】 陈玮；

【作者基本信息】 广东工业大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 机器人世界杯足球赛(Robot World Cup, Robocup),它涉及人工智能、机器人学、传感、通讯等诸多领域的前沿研究和技术集成。RoboCup2D仿真比赛系统作为一个仿真平台,在此基础上通过对Agent的动作技能及团队的协作策略等研究检验各种技术和智能算法理论。在动态、实时性、环境信息不完全确定的RoboCup平台中,各种因素影响着球队取得胜利,而选择哪些因素通过什么样的方法来解决问题,是研究的必要话题。本文以RoboCup2D为载体,从手工编码方式和机器学习方式两方面出发,研究学习综合评价法和机器学习算法,并围绕我们球队的Agent的高层技能动作展开研究。主要研究内容如下：首先,简要介绍了RoboCup的情况、研究现状及针对手工编码和机器学习的研究方式,然后运用这两种方式对Agent的高层动作技能进行研究。同时阐述了RoboCup2D仿真比赛平台以及与仿真平台相关的感知模型、运动模型和动作模型；分析得到GDUT-TiJi球队的整体结构和Agent结构模型。其次是对于射门技能的研究。通过分析球队底层代码以及观看比赛录像,总结了球队射门技能的不足以及影响射门成功的原因,选择采用手工编码的方式并运用评价的思想进行射门决策的设计。对比分析了目前主要的几种评价法之后,选择基于灰色关联度的灰色综合评价法建立机器人足球2D仿真比赛中Agent的射门技能决策。然后阐述了基于灰色关联度的灰色综合评价法的原理,建立了评价指标体系,设置了组合权重,并通过比赛实验提高了射门成功率,验证了此方法的可行性。最后,在传球技能的研究上,我们首先比较了手工编码方式和机器学习方式的优缺点,然后阐述了机器学习的意义。传球动作属于局部协作动作,适于用机器学习的方式进行研究。对比分析了机器学习的各种学习策略,结合RoboCup环境的特点,选用基于DFL的Agent自主学习方法进行研究。然后论文阐述了Agent自主学习概念、基于DFL的Agent心智模型结构及Agent自主学习模型,并给出一个实例验证了此方法的可行性和有效性。最后将此方法应用于球队中,选取防守策略强的队伍进行实验,在传球技能的学习上,取得了较好的效果。更多还原

【Abstract】 RoboCup World Cup Soccer Games and Conference(RoboCup) involves advanced research and novel technologies including artificial intelligence, robotics, sensing, communication and many other areas. Working as a simulation platform, RoboCup2D simulation game system is often employed to verify various intelligent algorithm theories and technologies by studying action skills of individual agent and collaboration strategy of the team. To emulate the constantly changing characteristic of soccer games,the RoboCup2D simulation game system offers a dynamic environment where only partial information is available to the players(agents). Under this circumstance, the choice of factors and decision-making strategies is critical to increase the winning ratio. Based on RoboCup2D platform, this thesis investigates methods of manual coding mode and machine learning for high-level actions of the agents in our team. The main work in this thesis consistes of three parts, as follows.Firstly, we make a brief introduction on the research status and existing methods involving RoboCup, with an emphasis on two major research paradigms, namely, manual coding mode and machine learning. These two paradigms are to be employed to study the decision-making of agent’s high-level action. The principle and models of sense, movement and action utilized by the soccer server are also described in this part. We also depict the structure and agent model in our team our team "GDUT-TiJi"Secondly, we focus on the study of shooting skill. After an analysis of our team’s code and game videos, we find the shortcomings of shooting skill and reasons resulting in shooting failures. We then select the manual coding way and use the grey comprehensive evaluation criterion to improve the shooting strategy according to characteristic of various comprehensive evaluation methods and RoboCup2D simulation game system. Experiment results suggest that this method improves the success rate of shooting.Finally, we investigate the passing skill. A comparison between the manual coding and machine learning leads to the conclusion that the latter is more suitable for decision-making process at the moment of passing the ball. Among various machine learning strategies, we choose the DFL-based autonomous learning as the main algorithm. After verifying the feasibility and validity of this method with an example, we finally apply this method in our team with an opponent team adopting a strong defensive strategy. Statistics shows that the rate of passing success is greatly increased by this method.更多还原