強(qiáng)化學(xué)習(xí)是目前熱門的研究方向。對(duì)不同強(qiáng)化學(xué)習(xí)的方法與paper進(jìn)行分類有助于我們進(jìn)一步了解針對(duì)不同的應(yīng)用場(chǎng)景，如何使用合適的強(qiáng)化學(xué)習(xí)方法。本文將對(duì)強(qiáng)化學(xué)習(xí)進(jìn)行分類并列出對(duì)應(yīng)的paper。
1. Model free RL
a. Deep Q-Learning系列
算法名稱：DQN
論文標(biāo)題：Playing Atari with Deep Reinforcement Learning
發(fā)表會(huì)議：NIPS Deep Learning Workshop, 2013.
論文鏈接：https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：5942
算法名稱：Deep Recurrent Q-Learning
論文標(biāo)題：Deep Recurrent Q-Learning for Partially Observable MDPs
發(fā)表會(huì)議：AAAI Fall Symposia, 2015
論文鏈接：https://arxiv.org/abs/1507.06527
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：877
算法名稱：Dueling DQN
論文標(biāo)題：Dueling Network Architectures for Deep Reinforcement Learning
發(fā)表會(huì)議：ICML, 2016
論文鏈接：https://arxiv.org/abs/1511.06581
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：1728
算法名稱：Double DQN
論文標(biāo)題：Deep Reinforcement Learning with Double Q-learning
發(fā)表會(huì)議：AAAI, 2016
論文鏈接：https://arxiv.org/abs/1509.06461
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：3213
算法名稱：Prioritized Experience Replay (PER)
論文標(biāo)題：Prioritized Experience Replay
發(fā)表會(huì)議：ICLR, 2016
論文鏈接：https://arxiv.org/abs/1511.05952
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：1914
算法名稱：Rainbow DQN
論文標(biāo)題：Rainbow: Combining Improvements in Deep Reinforcement Learning
發(fā)表會(huì)議：AAAI, 2018
論文鏈接：https://arxiv.org/abs/1710.02298
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：903
b. Policy Gradients系列
算法名稱：A3C
論文標(biāo)題：Asynchronous Methods for Deep Reinforcement Learning
發(fā)表會(huì)議：ICML, 2016
論文鏈接：https://arxiv.org/abs/1602.01783
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：4739
算法名稱：TRPO
論文標(biāo)題：Trust Region Policy Optimization
發(fā)表會(huì)議：ICML, 2015
論文鏈接：https://arxiv.org/abs/1502.05477
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：3357
算法名稱：GAE
論文標(biāo)題：High-Dimensional Continuous Control Using Generalized Advantage Estimation
發(fā)表會(huì)議：ICLR, 2016
論文鏈接：https://arxiv.org/abs/1506.02438
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：1264
算法名稱：PPO-Clip, PPO-Penalty
論文標(biāo)題：Proximal Policy Optimization Algorithms
發(fā)表會(huì)議：Arxiv
論文鏈接：https://arxiv.org/abs/1707.06347
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：4059
算法名稱：PPO-Penalty
論文標(biāo)題：Emergence of Locomotion Behaviours in Rich Environments
發(fā)表會(huì)議：Arxiv
論文鏈接：https://arxiv.org/abs/1707.02286
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：528
算法名稱：ACKTR
論文標(biāo)題：Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
發(fā)表會(huì)議：NIPS, 2017
論文鏈接：https://arxiv.org/abs/1708.05144
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：408
算法名稱：ACER
論文標(biāo)題：Sample Efficient Actor-Critic with Experience Replay
發(fā)表會(huì)議：ICLR, 2017
論文鏈接：https://arxiv.org/abs/1611.01224
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：486
算法名稱：SAC
論文標(biāo)題：Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor
發(fā)表會(huì)議：ICML, 2018
論文鏈接：https://arxiv.org/abs/1801.01290
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：1447
c. Deterministic Policy Gradients系列
算法名稱：DPG
論文標(biāo)題：Deterministic Policy Gradient Algorithms
發(fā)表會(huì)議：ICML, 2014
論文鏈接：http://proceedings.mlr.press/v32/silver14.pdf
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：1991
算法名稱：DDPG
論文標(biāo)題：Continuous Control With Deep Reinforcement Learning
發(fā)表會(huì)議：ICLR, 2016
論文鏈接：https://arxiv.org/abs/1509.02971
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：5539
算法名稱：TD3
論文標(biāo)題：Addressing Function Approximation Error in Actor-Critic Methods
發(fā)表會(huì)議：ICML, 2018
論文鏈接：https://arxiv.org/abs/1802.09477
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：839
d. Distributional RL系列
算法名稱：C51
論文標(biāo)題：A Distributional Perspective on Reinforcement Learning
發(fā)表會(huì)議：ICML, 2017
論文鏈接：https://arxiv.org/abs/1707.06887
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：600
算法名稱：QR-DQN
論文標(biāo)題：Distributional Reinforcement Learning with Quantile Regression
發(fā)表會(huì)議：AAAI, 2018
論文鏈接：https://arxiv.org/abs/1710.10044
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：188
算法名稱：IQN
論文標(biāo)題：Implicit Quantile Networks for Distributional Reinforcement Learning
發(fā)表會(huì)議：ICML, 2018
論文鏈接：https://arxiv.org/abs/1806.06923
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：139
算法名稱：Dopamine
論文標(biāo)題：Dopamine: A Research Framework for Deep Reinforcement Learning
發(fā)表會(huì)議：ICLR, 2019
論文鏈接：https://openreview.net/forum?id=ByG_3s09KX
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：107
e. Policy Gradients with Action-Dependent Baselines系列
算法名稱：Q-Prop
論文標(biāo)題：Q-Prop: Sample-Efficient Policy Gradient with An Off-Policy Critic
發(fā)表會(huì)議：ICLR, 2017
論文鏈接：https://arxiv.org/abs/1611.02247
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：259
算法名稱：Stein Control Variates
論文標(biāo)題：Action-depedent Control Variates for Policy Optimization via Stein’s Identity
發(fā)表會(huì)議：ICLR, 2018
論文鏈接：https://arxiv.org/abs/1710.11198
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：46
算法名稱：The Mirage of Action-Dependent Baselines in Reinforcement Learning
論文標(biāo)題：The Mirage of Action-Dependent Baselines in Reinforcement Learning
發(fā)表會(huì)議：ICML, 2018
論文鏈接：https://arxiv.org/abs/1802.10031
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：66
f. Path-Consistency Learning系列
算法名稱：PCL
論文標(biāo)題：Bridging the Gap Between Value and Policy Based Reinforcement Learning
發(fā)表會(huì)議：NIPS, 2017
論文鏈接：https://arxiv.org/abs/1702.08892
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：223
算法名稱：Trust-PCL
論文標(biāo)題：Trust-PCL: An Off-Policy Trust Region Method for Continuous Control
發(fā)表會(huì)議：ICLR, 2018
論文鏈接：https://arxiv.org/abs/1707.01891
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：68
g. Other Directions for Combining Policy-Learning and Q-Learning系列
算法名稱：PGQL
論文標(biāo)題：Combining Policy Gradient and Q-learning
發(fā)表會(huì)議：ICLR, 2017
論文鏈接：https://arxiv.org/abs/1611.01626
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：58
算法名稱：Reactor
論文標(biāo)題：The Reactor: A Fast and Sample-Efficient Actor-Critic Agent for Reinforcement Learning
發(fā)表會(huì)議：ICLR, 2018
論文鏈接：https://arxiv.org/abs/1704.04651
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：42
算法名稱：IPG
論文標(biāo)題：Interpolated Policy Gradient: Merging On-Policy and Off-Policy Gradient Estimation for Deep Reinforcement Learning
發(fā)表會(huì)議：NIPS, 2017
論文鏈接：http://papers.nips.cc/paper/6974-interpolated-policy-gradient-merging-on-policy-and-off-policy-gradient-estimation-for-deep-reinforcement-learning
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：117
算法名稱：Equivalence Between Policy Gradients and Soft Q-Learning
論文標(biāo)題：Equivalence Between Policy Gradients and Soft Q-Learning
發(fā)表會(huì)議：Arxiv
論文鏈接：https://arxiv.org/abs/1704.06440
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：170
h. Evolutionary Algorithms
算法名稱：ES
論文標(biāo)題：Evolution Strategies as a Scalable Alternative to Reinforcement Learning
發(fā)表會(huì)議：Arxiv
論文鏈接：https://arxiv.org/abs/1703.03864
當(dāng)前谷歌學(xué)術(shù)引用次數(shù)：802
