To reduce the training time for deep Q network(DQN), the paper researched on two classical control problems, cart pole and mountain car on Open AI Gym, by a DQN method combined with prioritized experience replay scheme and the dueling architecture(dueling DQN-PR). The prioritized experience replay was rank-based and adopted a deep neural network in the dueling architecture. The simulation results show that, compared with regular DQN, DQN with dueling network and DQN with prioritized experience replay, dueling DQN-PR acquires better learning performance with least training time. Meanwhile, the impacts of parameters on dueling DQN-PR were analyzed in detail, which provided valuable reference for the practical application.

[2020-02-05] Printed Article /doc/2018.06.0513/v2