Human-level control through deep reinforcement learning

[joyhuang9473]

Human-level control through deep reinforcement learning

• Author: Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg & Demis Hassabis
• Origin: https://www.nature.com/articles/nature14236
• Related:
  • https://sites.google.com/a/deepmind.com/dqn/home/Human_Level_Control_through_Deep_Reinforcement_Learning.zip?attredirects=0&d=1
  • https://chan-y-park.github.io/blog/rl_atari_part_1.html
  • https://github.com/chan-y-park/rl-atari
  • DQN从入门到放弃6 DQN的各种改进 https://zhuanlan.zhihu.com/p/21547911

[joyhuang9473]

250

[joyhuang9473]

DQN从入门到放弃6 DQN的各种改进 https://zhuanlan.zhihu.com/p/21547911

改进目标Q值计算：Deep Reinforcement Learning with Double Q-learning
改进随机采样：Prioritized Experience Replay
改进网络结构，评估单独动作价值：Dueling Network Architectures for Deep Reinforcement Learning ( 本文为ICML最佳论文之一）
改进探索状态空间方式：（1）Deep Exploration via Bootstrapped DQN （2）Incentivizing Exploration In Reinforcement Learning With Deep Predictive Models
改变网络结构，增加RNN：Deep Recurrent Q-Learning for Partially Observable MDPs（非DeepMind出品，效果很一般，谈不上改进，本文也不考虑讲解）
实现DQN训练的迁移学习：（1）Policy Distillation （2） Actor-Mimic: Deep Multitask and Transfer Reinforcement Learning
解决高难度游戏Montezuma‘s Revenge：Unifying Count-Based Exploration and Intrinsic Motivation
加快DQN训练速度：Asynchronous Methods for Deep Reinforcement Learning （这篇文章还引出了可以替代DQN的A3C算法，效果4倍Nature DQN）
改变DQN使之能够应用在连续控制上面：Continuous Deep Q-Learning with Model-based Acceleration