Edge-based Situ-aware Reinforcement Learning for Traffic Congestion Mitigation

Traffic congestion may cause elongated travel time, increased fuel consumption and extra pollution. To mitigate congestion, we propose a new approach based on multi-agent reinforcement learning (RL) to learn policies dictating path selections for vehicles. The algorithm utilizes the interactions between RL agents with Q-Learning and edge servers in monitoring traffic at road intersections. As an important difference between this work and existing approaches, we take human desire and realistic rewards into account. Extensive simulation experiments show that the resulting mechanism is promising and more RL agents can be incentive to follow rerouting directions when congestion is detected. Also, this algorithm has comparable performance as the Dynamic Dijkstra Algorithm.