This paper proposes an adaptive traffic light system for four-phase intersection using reinforcement learning. The existing adaptive traffic light system using two-phase order so far. While the four-phase order in the intersection can avoid vehicle conflict and decrease the vehicle waiting time. Therefore, the reinforcement learning algorithm is used to optimize the reward of reducing waiting time in the intersection through real-time traffic information and traffic light control formulation. The simulation results show the validation of reward can decrease the waiting time of the training system of deep neural networks (DNN).