Research And Implementation Of DTA Algorithm Based On FPGA Heterogeneous Platform

Dynamic Traffic Assignment(DTA)is a technology that utilizes real-time or predictive traffic information and demand to optimize and control traffic flow,aiming to alleviate congestion and improve road efficiency.With the advancement of technologies such as big data,edge computing,and the Internet of Things,the distributed model of DTA has been widely researched.Due to its characteristics of low complexity,high concurrency,and scalability,it is more suitable for the growing scale of urban traffic compared to centralized models.The combination of Multi-Agent Reinforcement Learning(MARL)algorithm and distributed DTA theory can further enhance the system’s dynamic adaptability,effectively coordinate the optimal solutions for both users and the system,and achieve rational allocation of traffic resources.However,the complex road network environment and time-varying traffic data present challenges to the storage and computing capabilities of distributed node devices.Designing a reasonable algorithm deployment plan and building an efficient hardware platform is a problem that requires a collaborative solution between software and hardware.To address these issues,this thesis improves the MARL algorithm by incorporating a traffic prediction module and graph decomposition of the state space.Additionally,the system is deployed on a vehicular FPGA heterogeneous platform using a distributed architecture,enabling realtime traffic allocation in large-scale road network environments.The specific tasks are outlined as follows:(1)In order to obtain future demand and traffic conditions for large-scale urban road networks,a traffic flow prediction method based on spatio-temporal graph attention network was studied.According to the traffic flow generation characteristics,the road traffic flow was divided into self-generated flow and multi-order neighbor diffusion flow.After extracting the temporal features,the self-generated flow decoder and diffusion graph attention network were respectively used to model the two types of flows,and the spatio-temporal features were fused to obtain the prediction results.Comparison experiments on multiple datasets show that the method can achieve optimal prediction performance while greatly reducing model parameter redundancy and computational resource consumption,enabling accurate prediction of large-scale urban traffic data,and providing macro-trend reference for subsequent MARL-based distributed DTA models.(2)To address the growing volume of data in large-scale urban traffic and balance the user equilibrium and system optimality criteria in DTA models while avoiding path convergence caused by personal navigation,a Multi-Agent Deep Deterministic Policy Gradient method with Traffic Prediction(TP-MADDPG)is studied.The traffic prediction module is integrated into the reinforcement learning model to enhance its estimation capability of future environmental state space,enabling long-term coordination in traffic allocation.To address the high dimensionality issue arising from a large number of agents in the joint state and action space,the road network graph is decomposed using multi-hop neighbor sampling to remove irrelevant state space information unrelated to the current agent’s itinerary,thus reconstructing the agent’s state space.A dual reward module is established to incentivize the agent’s action policy from both the system and user perspectives.The effectiveness of the proposed method is validated through simulation experiments on real-world datasets.(3)In terms of distributed deployment of the model,the training and inference process of the model is implemented based on an on-board FPGA heterogeneous platform to address the problem of long training time and limited resources of MADDPG in the case of large-scale intelligences.The hardware architecture is designed using techniques such as pipeline operation and brain floating point representation,and the neural network acceleration IP core is customized by High-level Synthesis(HLS)tools,and the ARM side is used to coordinate each component to achieve gas pedal invocation,model training,and cloud communication.On-board experiments show that the platform can achieve significant improvements in inference and training efficiency with lower power and fewer hardware resources.