DEEP INTELLIGENT NETWORK-DRIVEN MULTI-AGENT HIERARCHICAL REINFORCEMENT LEARNING FOR NEUROMORPHIC-ACCELERATED URBAN TRAFFIC FLOW OPTIMIZATION

Abstract

Urban traffic congestion continues to impose substantial economic and environmental costs worldwide. DIN-HRL is a Deep Intelligent Network–Hierarchical Reinforcement Learning framework that takes on three persistent weaknesses in existing DRL approaches: poor scalability, excessive energy consumption, and spatial representations that miss coordinated multi-intersection dynamics. The proposed system integrates directed hypergraph neural networks with spatio-temporal attention, a three-tier goal-conditioned policy hierarchy , and a calibrated ANN-to-SNN conversion pipeline for deployment on Intel Loihi 2 and SpiNNaker 2 neuromorphic platforms. Evaluated on a 64-intersection SUMO benchmark across five traffic scenarios, five random seeds, and nine state-of-the-art baselines, DIN-HRL achieved a throughput of 92.1 veh/h, representing a 77.1% improvement over fixed-time control , while reducing mean waiting time to 44.3 s and improving the safety score to 94.5. The neuromorphic implementation achieved the energy cost of 2.4 J/step, that is, 87.1% lower than with CPUs. Also, the latency of the implementation was 1.2 ms, 96.2% less than inference in GPU. A six-variant ablation study further confirms the contribution of each architectural component. Importantly, all hardware results were obtained through direct on-chip power-rail instrumentation rather than extrapolation from vendor specifications. DIN-HRL also converged 30.6% faster than MAPPO and retained 80% reward at 128 agents, establishing a strong benchmark for neuromorphic traffic signal control.