Deep Reinforcement Learning of the Model Fusion with Double Q-learning

Abstract

Q-learning algorithm is a one-step algorithm, which can overestimate the function of action value in some cases. Aiming at this problem, this paper uses double q - learning algorithm and the deep neural network to form deep Q network. Then combining with the different structure of deep neural network method to reduce sample correlation and increasing sample size in the experience bank. And it is tested in the Atari 2600 series video games, and it turns out that the method has the effect on some games.