【作者】 李岩；

【导师】 刘景泰；

【作者基本信息】 南开大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 随着移动计算、物联网等技术的迅速发展,网络机器人研究领域已经成为当前机器人研究中的热门研究课题,受到各国科研机构的高度重视。竞争型网络机器人系统以机器人之间的竞争关系为着眼点,与传统协作型机器人有着本质的区别。本文围绕竞争型网络机器人系统的关键问题以及相应解决方法展开了研究。竞争型网络机器人系统中存在的关键问题如下：一、二人零和对抗中的最优策略生成问题。在与对手智能体的作业目标是相互对立的情况下,如何制定最优策略才能够为己方带来最大利益是必须面对的问题。二、对手模型建立问题。建立有效的对手模型,才能够更加准确的估计对手的策略或意图等高层意识行为,为克敌制胜创造有利的信息条件。三、多源不确定性对系统的干扰问题。如何在具有多源不确定性的环境下进行感知、决策及行动是本系统的特有问题。本文围绕竞争型网络机器人系统的关键问题,得出如下研究结果：(1)针对二人零和对抗中的最优策略生成问题,本文引入了基于理性对手的策略生成方法。该方法通过微分对策将竞争型网络机器人系统的进攻与防守行为视为一种追逃博弈问题,通过找出双方的均衡策略,得出在双方都是理性决策者的情况下,双方的最优策略。本文通过仿真实验证明了在二人零和微分对策中,双方的最理性策略就是均衡策略。任何一方偏离其自身的最优策略都会使得对手收益增大,自己收益减小。(2)针对对手模型建立问题,本文应用了基于隐马尔可夫模型的对手意图识别方法。该方法通过将观察到的对手行为序列,离散化为对手攻防条件指数这一显状态,并将对手的攻防时机作为隐状态,分别训练在对手全力进攻、全力防守、普通防守、普通进攻四种意图下的隐马尔可夫模型。在对弈中,通过观察对手的进攻和防守行为所导致的对手攻防条件的变化,运用学习到的隐马尔可夫模型参数比较对手各种意图的概率值,其中最大的概率值对应的对手意图即为其当前意图。最后通过系统实验证明了本方法的有效性。(3)针对竞争型网络机器人系统的特点,提炼了竞争型网络机器人系统存在的多源不确定性,包括观测不确定性、执行不确定性、环境不确定性等。(4)针对多源不确定性对系统的干扰问题,整合了面向不确定性的竞争型网络机器人系统的分层体系结构。该体系结构分为策略层、执行层、物理接口层。其中策略层有两个作用,一是根据对手信息猜测对手意图,并根据对手历史行为分析对手水平；二是根据环境、对手、己方的状态,给出当前的态势评估结果,根据态势评估值选择相应策略。执行层的核心作用是在充满多源不确定性的环境下,根据策略层意图,选择期望收益最大的行动,从而有效克服不确定性对于系统的严重影响。物理接口层用于机器人获取外界信息,及提供针对当前机器人系统的轨迹规划等控制方法。本层的引入增强了本系统的可移植性及可扩展性。最后通过系统实验证明了该体系结构的优越性。更多还原

【Abstract】 With the rapid development of technologies such as mobile computing and Internet of Things, networked robot system has already become a hot subject in robotics field and has been highly valued by research institutions in every nation. The competitive networked robot system focuses on the competitive relationship between robots, which is fundamentally different from traditional collaborative robots. The dissertation focuses on the key issues in the competitive networked robot system and corresponding solutions.The key issues in the competitive networked robot system are as follows:Ⅰ. The generation of the optimal strategy in two-robot zero-sum game. When the goals between two intelligent agents are mutually contradictory, it is inevitable to generate the optimal strategy for each one’s own best interest.Ⅱ. Modeling of the opponent. The effective modeling of the opponent provides a more precise way to estimate the opponent’s high-level consciousness behavior, such as strategy and intention, therefore creates information advantage for victory.Ⅲ. The disturbing problems upon system deriving from multi-source uncertainties. The environment perception, decision-making and action under multi-source uncertainties are the peculiar problems in the competitive networked robot system.Concentrated on the key issues of competitive networked robot system, the dissertation draws the following results:(1) Aiming at the generation of the optimal strategy in two-robot zero-sum game, the dissertation introduces a strategy generation method based on rational opponent. The method, which is based on differential games, views the offensive and defensive behavior of competitive networked robot system as a pursuit game. Assuming that both sides are rational, this method finds the balanced strategy to generate the optimal strategy for both sides. The simulation results show that in a two-robot zero-sum differential game, the most rational strategy for both sides is balanced strategy. Anyone who deviates from its own optimal strategy will damage its own gains and benefit the opponent’s. (2) Regarding modeling the opponent, the dissertation applies a method of opponent’s intention recognition which is based on Hidden Markov Model (HMM). By taking the offensive and defensive condition index, which is discretized from the observed opponent’s behavior sequence, as explicit state and the opponent’s offensive and defensive time as hidden state, the method respectively trains four Hidden Markov Models when the opponent commits full attack, full defense, common attack and common defense. During a game, observe the change of the opponent’s offensive and defensive condition index, which derives from the opponent’s offensive and defensive behavior, and calculate the probability values based on the four HMM parameters above. At this time, the maximum probability value refers to the opponent’s intention. The validity of the method is verified with experiment results in the dissertaton.(3) According to the characteristics of competitive networked robot system, the dissertation refines the multi-source uncertainties existing in the system, including observation uncertainty, execution uncertainty and environmental uncertainty, etc.(4) Aiming at the disturbing problems upon system deriving from multi-source uncertainties, the dissertation proposes a hierarchy architecture for the competitive networked robot system. The architecture is divided into strategy layer, execution layer and the physical interface layer. One effect of the strategy layer is to recognize the opponent’s intention and analyze the rival level according to its history behaviors. The other one is to assess the current situation based on the circumstance, opponent and one’s own states, and to choose corresponding strategy according to the assessment. The execution layer is mainly used to cope with the multi-source uncertainties, and to select the action which may bring the maximum expected gains. The physical interface layer is used to gather information and to provide control method for current robot system such as trajectory planning. The introduction of the layer improves the portability and the extendibility of the system. Finally, the superiority of the architecture is verified with experiment results in the dissertation.更多还原