【作者】 魏善碧；

【导师】 柴毅；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 多智能体系统是控制理论和人工智能的一个前沿学科,其研究核心是寻求一种有效的方法使功能独立的智能体通过协作完成复杂的控制任务或解决复杂的问题。多智能体系统由多个空间分布的,具有独立动态特性智能体(子系统)组成,如多卫星系统、多机器人系统、多飞行器系统。智能体在网络化环境下相互协作共同完成系统级目标任务。智能体之间的协作主要体现在两个方面:与其它智能体拥有一个共同的目标;在实现目标的过程中在线共享信息。智能体具有的功能独立性(自治性)主要体现在能够独立地采集周围环境信息和其它智能体信息,并与其它智能体通信,计算并完成控制动作以达到自身的子目标。与其它控制方法相比,预测控制根据系统目标和工作环境变化,重新定义代价函数和约束条件,并使智能体在约束边界条件下工作,从而能够得到更好的控制效果,受到多智能体系统控制领域的青睐。然而集中式预测控制方法不能简单推广到分布式情形。在分布式控制中,智能体之间的代价函数耦合、耦合约束、有限通信条件约束(通信失效)等问题会引起智能体的独立控制行为发生冲突,影响智能体控制行为的一致性和系统全局稳定性。所以如何避免多智能体系统控制的冲突,提高控制行为的一致性和协作效率,成为多智能体系统控制的一个重要研究内容,在实际应用中有着重要的意义。本论文针对多智能体系统在全局代价函数耦合、耦合约束、通信失效情况下的控制问题开展了分布式预测方法研究。论文主要工作如下。针对代价函数耦合情况下的多智能体系统分布式控制问题,研究了基于相容约束和偏差惩罚的分布式预测控制方法。此方法分析了多智能体集中式全局代价函数的分解方法,得到了智能体分布式局部代价函数,构建了有限时域分布式预测控制问题,并为各智能体设计了不同的、时变的相容约束条件,在各智能体局部代价函数中加入设定状态轨迹和预测状态轨迹之间偏差函数项,对偏差进行惩罚,且在各采样时刻使各智能体的相容约束条件相对于前一时刻更强。与传统方法一样,采用性能代价函数作为李亚普洛夫函数分析了闭环稳定性。针对耦合约束情况下的多智能体系统分布式控制问题,研究了耦合约束条件的分布式预测控制方法。利用相容约束条件,将耦合约束处理为非耦合约束,并推广到避碰约束和通信距离约束等典型的耦合约束情况,分析了耦合约束成立的充分条件,证明了此方法的可行性和稳定性条件。针对多智能体系统复杂协调控制问题,研究了基于动态协调的混杂分布式预测控制方法,将智能体之间的避碰约束处理为基于方位的混杂规则,并在代价函数中引入布尔函数项。为适应复杂时变的协调控制问题,在每个采样时刻,根据各智能体当前位置和目标位置,设计动态协调规则以确定布尔函数项的权值,实现智能体之间的协调。针对带通信完全失效的多智能体系统分布式控制问题,研究了双模分布式预测控制方法。此方法将基于相容约束条件分布式预测控制方法和带邻居智能体状态优化的分布式预测控制方法相结合,在正常通信时,基于相容约束条件分布式预测控制方法中,智能体之间通过通信得到邻居智能体的设定状态轨迹;而在通信完全失效时,通过探测方式得到邻居智能体在预测时域的位置和速度信息,将基于邻居智能体状态的预测纳入智能体优化问题中,得到基于邻居智能体状态优化的分布式预测控制问题,从而实现带通信失效情况下的双模分布式预测控制问题。更多还原

【Abstract】 Multi-agent systems is the forum of control theory and artificial intelligence in which the central interest is to construct a kind of effective cooperation mechanism which enables function independent agent to achieve a complex controlling task, or solve a complex problem. A multi-agent system is composed of many spatial distributed subsystems with decoupled dynamics, such as multi-satellite systems, multi-robot systems, multi-aircraft systems, etc. Under the network environment, the system-level control objective is achieved by cooperation among agents. Cooperation refers to the agreement among the agents: Each agent has a common objective with neighbor agents, and share information online to realize the objective. Agents are function independent (autonomous) in that they are individually capable of sensing their environment and possibly other agents, communicating with other agents, and computing and implementing control actions to meet their portion of the objective.Compared to the traditional approach, model predictive control has the ability to redefine cost functions and constraints as needed to reflect changes in the system and/or the environment. Hence, MPC is extensively applied to the cooperative control of multi-agent systems, which makes the agents operate close to the constraint boundaries, and obtain better performance than traditional approaches. However, the research results of centralized model predictive control could not be simply extended to distributed model predictive control. For distributed control problem, the coupled global cost function, coupled constraints, limited communication constraints (communication failure) etc, need to be considered, which could lead to conflictions among the independent control actions, affect the consistency of control actions and global stability of system. Hence, how to avoid conflictions among agents, or improve the consistency and efficiency, becomes an important issue in the area of multi-agent systems, which is very important for engineering applications.So this thesis focuses on distributed predictive control (DMPC) for multi-agent systems under the coupled global cost function, coupled constraint, limited communication constraint (complete communication failure). The main work is as follows.For the distributed control problem of multi-agent systems with the coupled global cost function, this paper addresses an improved DMPC scheme using deviation penalty and compatibility constraint. This scheme analyzes the decomposition approach for the global and centralized cost function, gets the local and distributed cost function of each agent, establishes the distributed predictive control problem with finite horizon, and designs different and time-varying compatibility constraint for each agent. The deviation between what an agent is actually doing and what its neighbors believe that agent is doing is penalized in the local cost function of each agent. At each sampling instant the compatibility constraint of each agent is set tighter than the previous sampling instant. Like the traditional approach, the performance cost is utilized as the Lyapunov function to prove closed-looped stability.For the distributed control problem of multi-agent systems with coupled constraints, this paper addresses an improved DMPC scheme. The coupled constraints are transformed as non-coupled constraints. The results are extended to the typical coupling constraints such as avoidance constraint and limited distance constraint. The sufficient condition for coupled constraints is analyzed. Furthermore, the feasibility and stability of this method are investigated.For complex cooperative control problem of multi-agent systems, distributed model predictive control scheme based on the dynamic cooperative rules is addressed. The collision avoidance constraints are transformed as hybrid rules based on the positions, and the Boolean function term is incorporated in the cost function. In order to accommodate the complex time-varying environment, at each sample instant, the dynamic cooperative rules are designed according to the relative positions between the agents and between each agent and its destination, so as to determine the weights in the Boolean function.For the distributed control problem of multi-agent systems with complete communication failure, this paper proposes a dual-mode DMPC scheme. This scheme combines the DMPC with compatibility constraint and the DMPC with neighbor agent state optimization. For the normal communication, by using the DMPC with compatibility constraint, each agent gets the assumed state trajectory of its neighbors through communication. For the complete communication failure, by using DMPC with neighbor agent state optimization, each agent only detects the local information (position and velocity), including the prediction of neignbos agent’s state into its optimal problem. Each agent constructs the DMPC schme based on the optimization of neighbor agent state, realizes the dual-mode DMPC of multi-agent systems with complete communication failure .更多还原