Urban traffic congestion is a multi-intersection coordination problem: optimising one junction in isolation can shift delays elsewhere. This project develops cooperative multi-agent deep reinforcement learning (MARL) controllers that jointly optimise throughput across networks of intersections.
Research has progressed from single-intersection baseline agents through to multi-agent systems that communicate implicit state representations, operating in real time from live camera feeds rather than ground-truth simulation data. Experiments on the Traffic3D platform demonstrate consistent reductions in average waiting time and emissions compared to fixed-cycle and single-agent baselines.