毕业生知识库

    城市交通系统作为一个典型的社会物理信息系统（Cyber-Physical-Social Systems，CPSS），其研究面临着工程技术与社会动态的双重复杂性。这种复杂性不仅给城市交通系统的建模研究和实验分析造成了困难，也对交通管理的智能化水平提出了更高要求。在此背景下，平行交通系统的理念应运而生。该理念基于ACP（Artificial societies, Computational experiments, and Parallel execution）方法，通过构建与实际交通系统相对应的人工系统，并采用虚实互动与平行执行的方式，实现了对实际交通系统的优化管理和引导控制，为智能交通领域提供了新的研究思路和实践方向。

    基础模型作为 ACP中人工社会的具体实现，通过大规模计算实验挖掘现实世界的隐性知识，进而促进了基础智能的涌现和泛化。这一过程为基于CPSS的平行交通系统研究提供了新的视角。本文面向平行交通系统的大规模路网信号控制问题，展开了城市交通控制基础模型关键技术研究。研究内容主要包括以下几个方面：首先，基于人工交通系统，研究了大规模路网协同控制方法，以满足模型构建的基本功能需求；其次，基于虚实融合的多样化交通场景，研究了泛化性增强方法，为模型在不同交通场景中的迁移和应用提供了基础；最后，针对实际交通场景的自适应控制需求，研究了提升模型对动态场景响应性的优化策略，以增强模型在实际交通场景的适应能力和控制效果。本文的主要工作如下：

1. 设计了一种基于基础模型的平行交通控制框架。该框架通过构建交通控制基础模型，为人工交通系统与实际交通系统的交互提供一种新方法，从而增强平行交通控制中的虚实交互过程。结合平行学习理论，该模型能够在人工系统中自我优化，为实际交通管理提供策略性指导。为了更好地推动模型优化及其在实际中的应用，提出了一种云边协同控制流程，该流程旨在充分利用模型的泛化能力，并满足交通控制系统对实时响应的需求。此外，为了支持这一流程的技术研究，构建了TengYun计算实验平台，提供了多样化的交互式环境，以便于模型的训练和性能评估。
2. 提出了一种融合全局动态表征的分布式控制方法，以提升大规模路网的协同控制性能。该方法结合了云端的Transformer表征模块和本地端的强化学习决策模块，分别负责全局信息建模和局部控制策略的实施。通过这两个模块的平行执行与协同优化，该方法整合了集中式控制的动态建模能力和分布式控制的灵活性，从而提升模型在协同控制方面的性能。此外，为了更准确地对交通路网信息进行动态建模，设计了一种适用于交通信号控制任务的相对位置编码。该方案有助于模型更准确地模拟交通信息的动态传播过程并进行合理的注意力机制分配，从而提高对关键交通动态的识别和响应能力。在基于杭州和济南路网真实数据构建的人工交通系统中进行的计算实验结果表明，本方法在多种场景中的协同控制性能优于现有的（State of the Art, SOTA）方法。
3. 提出了一种虚实混合场景下的多场景联合学习方法，以增强模型在不同交通场景中的泛化能力。该方法将交通控制问题建模为多场景多智能体马尔可夫决策过程，扩展了模型的状态探索空间。通过协同优化多个场景的决策过程，模型能够提取跨场景的共性特征，并建立从状态空间到动作空间的鲁棒映射关系，从而增强模型对新场景的适应性和泛化能力。针对真实场景训练数据不足的问题，提出了一种基于Wasserstein生成对抗网络与梯度惩罚的人工交通场景构建方法，该方法能够利用有限的真实数据生成多样化的人工交通场景，为模型训练提供丰富的数据支持。在杭州路网的虚实融合场景数据集上的计算实验结果表明，本方法在鲁棒性和迁移性方面的表现优于现有的SOTA算法。
4. 提出了一种基于提示生成的分层控制方法，以优化模型面向动态交通场景的响应能力。该方法将交通控制问题建模为半马尔可夫决策过程，增强了模型处理时间序列决策的抽象表示能力，以适应交通模式的动态变化。采用Option-Critic算法，实现了云端与边缘端分层模型的协同优化：云端模型负责分析全局交通状况，并自动生成关于长期交通模式的决策提示；本地模型则依据这些提示及本路口的实时数据，执行短期交通扰动的自适应控制。为了提升云端模型的提示生成能力，引入了一种基于图Transformer的交通动态表征方法，使模型能够更有效地处理交通场景中的非欧几里得结构数据，并为本地模型提供多尺度信息支持。在杭州路网全天动态场景数据集上的计算实验结果表明，本方法在本地路口的自适应控制性能优于现有的先进算法，并在云端管控的提示决策定性分析中展现出良好特性。

       Urban transportation systems, as typical Cyber-Physical-Social Systems (CPSS), face the dual complexities of engineering technology and social dynamics. These complexities pose significant challenges to system modeling and experimental analysis, thereby demanding higher levels of transportation intelligence. In response, Parallel Transportation Systems (PTS) have been developed, rooted in the ACP (Artificial societies, Computational experiments, and Parallel execution) methodology. It involves constructing Artificial Transportation Systems (ATS) that mirror the actual systems, employing virtual-real interaction and parallel execution to optimize management and control strategies. This framework provides new insights and practical directions for advancing intelligent transportation research and practice.

       Foundation models, as concrete implementations of artificial societies within the ACP framework, utilize large-scale computational experiments to uncover implicit knowledge from the real world. This process fosters the emergence and generalization of foundational intelligence, offering a novel perspective for exploring CPSS-based PTS. This dissertation focuses on the key technologies of the Traffic Control Foundation Model (TCFM) within PTS, focusing on large-scale urban network signal control challenges. The research encompasses several key aspects: First, it investigates coordination control methods for large-scale road networks based on ATS to meet the basic functional requirements of the model. Second, it explores strategies to enhance model generalization in diverse traffic scenarios, ensuring the model's adaptability across various traffic environments. Finally, it focuses on optimization strategies that enhance the model's responsiveness to dynamic traffic scenarios, improving its adaptability and control effectiveness in real traffic conditions. The contributions of this dissertation are as follows:

1. Design of a parallel traffic control framework based on foundation models. This framework establishes a TCFM to facilitate interactions between artificial and actual transportation systems, enhancing the virtual-real interaction process in parallel traffic control. Based on the parallel learning theory, this model self-optimizes within the artificial systems, providing guidance for actual traffic management. To further enhance model optimization and its real-world applications, a cloud-edge collaborative control workflow is proposed to fully leverage the model's generalization capabilities and meet the real-time response requirements of traffic control systems. The TengYun computational experimental platform has been developed to support this process, offering a diverse interactive environment for model training and performance evaluation.
2. Development of a distributed control method with global dynamic representation. This method enhances the coordinated control performance of large-scale road networks by synergizing a cloud-based Transformer representation module with a local reinforcement learning decision-making module. Through parallel execution and collaborative optimization, this approach integrates the dynamic modeling capabilities of centralized control with the flexibility of distributed control, significantly improving the model's performance in coordinated control scenarios. Additionally, a relative position encoding mechanism is proposed for traffic signal control tasks, enabling the model to simulate the propagation of traffic information more accurately and allocate attention mechanisms effectively. Computational experiments using data from the Hangzhou and Jinan road networks demonstrate that this method outperforms existing State of the Art (SOTA) methods in various scenarios.
3. Proposal of a multi-scenario joint learning method for hybrid virtual-physical scenarios. This method models the traffic control problem as a multi-scenario multi-agent Markov Decision Process, expanding the model's state exploration space. By optimizing decision-making collaboratively across multiple scenarios, the model extracts common features and establishes robust mappings from state spaces to action spaces, enhancing its adaptability and generalization in new scenarios. To address the shortage of real-scenario training data, a method for constructing artificial traffic scenes based on Wasserstein Generative Adversarial Networks (WGAN) with Gradient Penalty is introduced. This method generates a diverse range of artificial traffic scenes from limited real data, enriching the dataset for model training. Computational experiments on a hybrid dataset of Hangzhou's road network show that this method surpasses some existing SOTA algorithms in robustness and transferability.
4. Introduction of a prompt-based hierarchical control method. This method models the traffic control problem as a semi-Markov Decision Process, enhancing the model's capability for abstract representation in time-series decisions and adapting to dynamic traffic patterns. It facilitates collaborative optimization between cloud and edge-level models within a hierarchical framework. The cloud model analyzes overall traffic conditions and generates decision prompts for long-term traffic patterns, while the local model adapts to short-term disturbances based on these prompts and real-time data at intersections. To enhance the cloud model's prompt generation capabilities, a Graph Transformer-based method for representing traffic dynamics has been introduced, enabling the model to effectively process non-Euclidean structured traffic data and provide multi-scale information support to local models. Computational experiments on Hangzhou's dynamic road dataset reveal that our method surpasses existing algorithms in local intersection adaptive control and demonstrates robust qualitative traits in cloud-managed traffic decision-making.