Field Implementation and Operational Validation of a Decentralized Graph-Based Multi-Agent Reinforcement Learning Traffic-Signal Controller on a Smart Corridor

Abstract

Adaptive traffic-signal control can respond to stochastic and directionally asymmetric demand, but transferring reinforcement-learning controllers from simulation to live infrastructure requires integration with real sensing, controller timing, pedestrian service, communication, and actuation constraints. This paper presents the field implementation and operational validation of a previously developed decentralized graph-based multi-agent reinforcement learning (DGMARL) traffic-signal controller on Chattanooga’s MLK Smart Corridor. The trained policy remained fixed during deployment and operated in inference-only mode without online learning. Field observations from computer vision, signal phase and timing (SPaT), controller-referenced timing, and neighboring agents were mapped into the policy-facing state, while live pedestrian requests and controller constraints were handled by deterministic execution logic. Policy-generated actions were validated before controller-facing signal states were transferred through an NTCIP-facing interface. Three supervised field trials across four intersections demonstrated end-to-end integration, concurrent multi-intersection execution, controller-referenced timing, configured phase and clearance progression, and retention and subsequent service of recorded pedestrian requests. For retained Trial 1 records, field-loop input-to-output timing indicators had median values of approximately 30–47 ms and 95th-percentile values of approximately 97–106 ms. The trials evaluate operational integration rather than comparative traffic-performance improvement. The results provide empirical evidence that a previously trained DGMARL framework can be transferred from a real-data-informed Digital Twin environment to live smart-corridor infrastructure while preserving the learned decision structure and accommodating field-specific sensing, timing, and controller requirements.