Abstract:Different from the traditional deep reinforcement learning method of training through transitions selected one by one from the experience replay, for the Deep Q Network (DQN) that uses the entire episode trajectory as the training sample, a method for expanding episode samples is proposed, which is based on genetic algorithm crossover operators. The episode trajectory is generated during the trial-and-error decision-making process of the interaction between the agent and the environment, in which similar key states will be encountered. With the similar state in the two episode trajectories as the intersection point, the episode trajectory that has not appeared till present can be generated to enlarge the number of episode samples and increase their diversity, thereby enhancing the agent’s exploration ability and improving sample efficiency. Compared with DQN that randomly selects samples and uses the Episodic Backward Update (EBU) algorithm, the proposed method can achieve higher rewards in the Playing Atari 2600.