摘要：Advances in deep reinforcement learning have allowed autonomous agents to perform well on video games, often outperforming humans, using only raw pixels to make their decisions. However, timely context awareness is not fully integrated. In this paper, we extend Deep Q-network (DQN) with spatio-temporal architecture - a novel framework that handles the temporal limitation problem. To incorporate spatio-temporal information, we construct variants of architectures by feeding spatial and temporal representations into Deep Q-networks in different ways, which are DQN with convolutional neural network (DQN-Conv), DQN with LSTM recurrent neural network (DQN-LSTM), DQN with 3D convolutional neural network (DQN-3DConv), and DQN with spatial and temporal fusion (DQN-Fusion), to explore the mutual but also fuzzy relationship between them. Extensive experiments are conducted on popular mobile game Flappy Bird and our framework achieves superior results when compared to baseline models.
© 2020, Springer Nature Switzerland AG.

ISSN号：0302-9743

卷、期、页：v 12533 LNCS,p283-294

发表日期：2020-01-01

期刊分区（SCI为中科院分区）：无

收录情况：EI（工程索引）

发表期刊名称：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

参与作者：李卫民

通讯作者：刘泽宇

第一作者：刘建伟,左信

论文类型：会议论文

论文概要：刘泽宇,刘建伟,李卫民,左信，Deep Reinforcement Learning with Temporal-Awareness Network，Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)，2020，v 12533 LNCS,p283-294

论文题目：Deep Reinforcement Learning with Temporal-Awareness Network