Urban Intersection Simulation and Verification via Deep Reinforcement Learning Algorithms

Abstract Reinforcement Learning (RL) uses rewards to have iteration and update the next state for training in an unknown complex environment. This paper aims find a possible solution traffic congestion problem train four Deep (DRL) algorithms verify urban intersection simulation environment different discussed dimensions, including practicability, efficiency, safety, complexity, limitation. The experiment result shows that DRL are efficient RL simulation. has succeeded verifying this comparison expands with three conclusions: agent can by Q-network(DQN), DoubleDQN, DuelingNet DQN, Categorical DQN be practical efficient. As results show, takes less time finish training, reduced collision rate after while. However, lacks causing limitations solving more problems, lack of pedestrian behaviors prediction emergency events. recommends creating includes exceptional cases pressure improve faster safer response future automatic driving.