The invention discloses an EADP controller off-line training method, an EADP controller off-line training system, an EADP controller online training method and an EADP controller online training system, wherein the methods and the systems are used for controlling a traffic signal of an intersection. The EADP controller off-line training method comprises the steps of according to an obtained system state and a constructed Action network and a constructed Critic network of each sub-ADP controller, defining a reward function, a system control parameter and a performance index; alternately training the Critic network of each sub-ADP controller according to the performance index and the reward function and alternately training the Action network of each sub-ADP controller according to the performance index and the reward function, thereby updating the weight of the Critic network and the weight of the Action network; and a fact that training reaches a training object is determined, recording the weight of the action network and the weight of the Critic network of each sub-ADP controller. Through the EADP controller offline training method, a technical problem of high difficulty for ensuring stability of a traditional ADP controller is settled, and furthermore self-adaptive control to the traffic signal is realized.