论文部分内容阅读
交通信号控制系统在物理位置和控制逻辑上分散于动态变化的网络交通环境，将每个路口的交通信号控制器看做一个异质的智能体，非常适合采用“无模型、自学习、数据驱动”的多智能体强化学习方法建模与描述。为了解该方法的研究现状、存在问题及发展前景，系统跟踪了多智能体强化学习在国内外交通控制领域的具体应用，包括交通信号MARL控制概念模型、完全孤立的多智能体强化学习控制、部分状态合作的多智能体强化学习控制和动作联动的多智能体强化学习控制，分析其技术特征和代际差异，讨论了多智体强化学习方法在交通信号控制中的研究动向，提出了发展网络交通信号多智能体强化学习集成控制的关键问题在于强化学习控制机理、联动协调性、交通状态特征抽取和多模式整合控制。
Traffic signal control system is physically and logically dispersed in a dynamic network traffic environment, and the traffic signal controller at each intersection is considered as a heterogeneous agent, which is very suitable for adopting "model-less, self-learning and data-driven Multi-agent reinforcement learning method modeling and description. In order to understand the research status, existing problems and development prospects of this method, the system systematically traced the specific applications of multi-agent reinforcement learning in the field of traffic control at home and abroad, including conceptual model of traffic signal MARL control, completely isolated multi-agent enhanced learning control, Partially state-agnostic multi-agent reinforcement learning control and action-linked multi-agent reinforcement learning control, analysis of their technical characteristics and intergenerational differences, discussed the research trend of multi-agent reinforcement learning method in traffic signal control, and proposed the development Network Traffic Signal Multi-Agent Intensive Learning Integrative control of the key issues is to strengthen the learning control mechanism, linkage coordination, traffic status feature extraction and multi-mode integrated control.