Reinforcement Learning for Isolated Intersection Signal Control

Fixed timing cannot adapt to demand fluctuation, while adaptive systems like SCATS rely on extensive threshold tuning. RL learns policies by interacting with simulation, but reward design and training stability remain hard.

Explore the feasibility of RL for isolated intersection signal control and its adaptability to demand fluctuation.

State: queue lengths of all approaches & elapsed phase time; action: 4-phase choice; reward: negative weighted delay plus switch penalty. Trained DQN in SUMO+TraCI for 2000 episodes with replay buffer & target network.

RL performs well in simulation, but the sim-to-real gap is large. Next I will use domain randomization and calibrate demand with real FCD to narrow the transfer gap.