18 Issues in Current Deep Reinforcement Learning from ZhiHu

  • 时间:
  • 浏览:1

强度强化学习的大问题在哪里?未来为啥走?或多或少方面可不都可不上能突破?

现有解法依然是蒙特卡洛搜索,详情可不都可不上能参考初代AlphaGo的实现【Silver et al 2016a】

under-appreciated reward exploration 【Nachum et al 2017)】

现有解法:

Duel DQN【Wang 2016c】(ICML2016最佳论文)

model-free与model-based的结合使用【Chebotar et al 2017】

Baker, B., Gupta, O., Naik, N., and Raskar, R. (2017). Designing neural network architectures using reinforcement learning. In the International Conference on Learning Representations (ICLR).

目前解法有一二个流派,一图胜千言:

Sutton, R. S., Mahmood, A. R., and White, M. (2016). An emphatic approach to the problem of off-policy temporal-difference learning. The Journal of Machine Learning Research, 17:1–29

Jaderberg, M., Mnih, V., Czarnecki, W., Schaul, T., Leibo, J. Z., Silver, D., and Kavukcuoglu, K. (2017). Reinforcement learning with unsupervised auxiliary tasks. In the International Conference on Learning Representations (ICLR).

Mahmood, A. R., van Hasselt, H., and Sutton, R. S. (2014). Weighted importance sampling for off-policy learning with linear function approximation. In the Annual Conference on Neural Information Processing Systems (NIPS).

Zhu, X. and Goldberg, A. B. (809). Introduction to semi-supervised learning. Morgan & Claypool

integrate temporal abstraction with intrinsic motivation 【Kulkarni et al 2016】

Watkins, C. J. C. H. and Dayan, P. (1992). Q-learning. Machine Learning, 8:279–292

Gu, S., Lillicrap, T., Ghahramani, Z., Turner, R. E., and Levine, S. (2017). Q-Prop: Sampleefficient policy gradient with an off-policy critic. In the International

Conference on Learning Representations (ICLR).

现有解法是Guided Policy Search 【Levine et al 2016a】

Chebotar, Y., Hausman, K., Zhang, M., Sukhatme, G., Schaal, S., and Levine, S. (2017). Combining model-based and model-free updates for trajectory-centric reinforcement learning. In the International Conference on Machine Learning (ICML)

Ho, J. and Ermon, S. (2016). Generative adversarial imitation learning. In the Annual Conference on Neural Information Processing Systems (NIPS).

learn, plan, and represent knowledge with spatio-temporal abstraction at multiple levels

O'Donoghue, B., Munos, R., Kavukcuoglu, K., and Mnih, V. (2017). PGQ: Combining policy gradient and q-learning. In the International Conference on Learning Representations (ICLR).

Anschel, O., Baram, N., and Shimkin, N. (2017). Averaged-DQN: Variance reduction and stabilization for deep reinforcement learning. In the International Conference on Machine Learning (ICML).

Instability and Divergence when combining off-policy,function approximation,bootstrapping

Li, K. and Malik, J. (2017). Learning to optimize. In the International Conference on Learning Representations (ICLR).

Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2016). Prioritized experience replay. In the International Conference on Learning Representations (ICLR).

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention Is All You Need. ArXiv e-prints.

Williams, R. J. (1992). Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3):229–256.

Nachum, O., Norouzi, M., Xu, K., and Schuurmans, D. (2017). Bridging the Gap Between Value and Policy Based Reinforcement Learning. ArXive-prints.

Distributed Proximal Policy Optimization 【Heess 2017】

Heess, N., TB, D., Sriram, S., Lemmon, J., Merel, J., Wayne, G., Tassa, Y., Erez, T., Wang, Z., Eslami, A., Riedmiller, M., and Silver, D. (2017). Emergence of Locomotion Behaviours in Rich Environments. ArXiv e-prints

PGQ,policy gradient and Q-learning 【O'Donoghue et al 2017】

Sutton, R. S. (1990). Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In the International Conference on Machine Learning (ICML).

Finn, C., Christiano, P., Abbeel, P., and Levine, S. (2016a). A connection between GANs, inverse reinforcement learning, and energy-based models. In NIPS 2016 Workshop

on Adversarial Training.

data storage over long time, separating from computation

现有解法完整性围绕迁移学习走 【Taylor and Stone, 809、Pan and Yang 2010、Weiss et al 2016】,learn invariant features to transfer skills 【Gupta et al 2017】

Emphatic-TD 【Sutton 2016】

现有解法基本上是learn to learn

Tessler, C., Givony, S., Zahavy, T., Mankowitz, D. J., and Mannor, S. (2017). A deep hierarchical approach to lifelong learning in minecraft. In the AAAI Conference on Artificial Intelligence (AAAI).

one/few/zero-shot learning 【Duan et al 2017、Johnson et al 2016、 Kaiser et al 2017b、Koch et al 2015、Lake et al 2015、Li and Malik 2017、Ravi and Larochelle, 2017、Vinyals et al 2016)

万变不离其宗,Temporal Difference土最好的办法仍然是策略评估的核心哲学【Sutton 1988】。TD的拓展版本和她本身一样鼎鼎大名——1992年的Q-learning与2015年的DQN。

Sutton老爷子教科书里的经典安利:Dyna-Q 【Sutton 1990】

Audiffren, J., Valko, M., Lazaric, A., and Ghavamzadeh, M. (2015). Maximum entropy semisupervised inverse reinforcement learning. In the International Joint Conference on Artificial Intelligence (IJCAI).

现有解法基本上围绕模仿学习

lifelong learning with hierarchical RL 【Tessler et al 2017】

Bahdanau, D., Brakel, P., Xu, K., Goyal, A., Lowe, R., Pineau, J., Courville, A., and Bengio, Y. (2017). An actor-critic algorithm for sequence prediction. In the International

Conference on Learning Representations (ICLR).

exploration-exploitation tradeoff

prediction, policy evaluation

Horde 【Sutton et al 2011】

这里精选18个关键大问题,蕴含空间搜索、探索利用、策略评估、内存使用、网络设计、反馈激励等等话题。本文精选了73篇论文(其中2017年论文有27篇,2016年论文有21篇)为了方便阅读,原标题倒进文章最后,可不都可不上能根据索引找到。

Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwinska, A., Col- ´ menarejo, S. G., Grefenstette, E., Ramalho, T., Agapiou, J., nech Badia, A. P., Hermann, K. M., Zwols, Y., Ostrovski, G., Cain, A., King, H., Summerfield, C., Blunsom, P., Kavukcuoglu, K., and Hassabis, D. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538:471–476

Osband, I., Blundell, C., Pritzel, A., and Roy, B. V. (2016). Deep exploration via bootstrapped DQN. In the Annual Conference on Neural Information Processing Systems (NIPS).

美中缺陷,TD Learning中很容易跳出Over-Estimate(高估)大问题,具体意味着着 如下:

learn with MDPs both with and without reward functions 【Finn et al 2017)】

Schulman, J., Abbeel, P., and Chen, X. (2017). Equivalence Between Policy Gradients and Soft Q-Learning. ArXiv e-prints.

Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., & Tassa, Y., et al. (2015). Continuous control with deep reinforcement learning. Computer Science, 8(6), A187.

Xu, K., Ba, J. L., Kiros, R., Cho, K., Courville, A.,Salakhutdinov, R., Zemel, R. S., and Bengio,Y. (2015). Show, attend and tell: Neural image caption generation with visual attention. In the International Conference on Machine Learning (ICML).

这5天我阅读了两篇篇猛文A Brief Survey of Deep Reinforcement Learning 和 Deep Reinforcement Learning: An Overview ,作者排山倒海的引用了80多篇文献,阐述强化学习未来的方向。原文归纳出强度强化学习中的常见科学大问题,并列出了目前解法与相关综述,我在这里做出分类整理,抽取了相关的论文。

Mnih, V., Badia, A. P., Mirza, M., Graves, A., Harley, T., Lillicrap, T. P., Silver, D., and Kavukcuoglu, K. (2016). Asynchronous methods for deep reinforcement learning. In the International Conference on Machine Learning (ICML)

learn with expert's trajectories and those may not from experts 【Audiffren et al 2015】

variational information maximizing exploration 【Houthooft et al 2016】

TODO list:文章内容还缺陷充实,或多或少论文是全的。未来一段时间会把论文的链接找齐,下载好或多或少打个包传到百度云上,预计一5天完成。(2017/12/19)

benefit from non-reward training signals in environments

learn from demonstration 【Hester et al 2017】

现有解法完整性围绕半监督学习 【Zhu and Goldberg 809】

unify count-based exploration and intrinsic motivation 【Bellemare et al 2017】

train dialogue policy jointly with reward model 【Su et al 2016b】

learn to navigate with unsupervised auxiliary learning 【Mirowski et al 2017】

(neural networks architecture design )

Pan, S. J. and Yang, Q. (2010). A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345 – 1359.

Nachum, O., Norouzi, M., and Schuurmans, D. (2017). Improving policy gradient by exploring under-appreciated rewards. In the International Conference on Learning Representations (ICLR).

Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A.,

Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. (2015).

Human-level control through deep reinforcement learning. Nature, 518(7540):529–533.

吴恩达的逆强化学习【Ng and Russell 800)】

Barto, A. G. and Mahadevan, S. (803). Recent advances in hierarchical reinforcement learning. Discrete Event Dynamic Systems, 13(4):341–379.

现有解法:多层强化学习 【Barto and Mahadevan 803】

Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H., Leibo, J. Z., Munos, R., Blundell, C., Kumaran, D., and Botvinick, M. (2016a). Learning to reinforcement learn. arXiv:1611.05763v1.

learning to learn, 【Duan et al 2017、Wang et al 2016a、Lake et al 2015】

learn knowledge from different domains

van Hasselt, H., Guez, A., , and Silver, D. (2016a). Deep reinforcement learning with double Qlearning. In the AAAI Conference on Artificial Intelligence (AAAI).

异步算法A3C 【Mnih 2016】

stochastic neural networks for hierarchical RL 【Florensa et al 2017】

Q-learning与Actor-Critic

control, finding optimal policy

Kaiser, Ł., Nachum, O., Roy, A., and Bengio, S. (2017b). Learning to Remember Rare Events. In the International Conference on Learning Representations (ICLR).

Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., and Wierstra, D. (2016). Matching networks for one shot learning. In the Annual Conference on Neural Information Processing Systems (NIPS).

Levine, S., Finn, C., Darrell, T., and Abbeel, P. (2016a). End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17:1–40.

最出名的解法是在Nature上大秀一把的Differentiable Neural Computer【Graves et al 2016】

Mirowski, P., Pascanu, R., Viola, F., Soyer, H., Ballard, A., Banino, A., Denil, M., Goroshin, R., Sifre, L., Kavukcuoglu, K., Kumaran, D., and Hadsell, R. (2017). Learning to navigate in complex environments. In the International Conference on Learning Representations (ICLR).

Sutton, R. S. and Barto, A. G. (2017). Reinforcement Learning: An Introduction (2nd Edition, in preparation). MIT Press.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., , and Bengio, Y. (2014). Generative adversarial nets. In the Annual

Conference on Neural Information Processing Systems (NIPS), page 2672?2680.

van Hasselt, H. (2010). Double Q-learning. Advances in Neural Information Processing Systems 23:, Conference on Neural Information Processing Systems 2010.

下面几篇论文都是DQN相关话题的:

adapt rapidly to new tasks

train perception and control jointly end-to-end

Sutton, R. S., Modayil, J., Delp, M., Degris, T., Pilarski, P. M., White, A., and Precup, D. (2011). Horde: A scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction, , proc. of 10th. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS).

经验回放下的actor-critic 【Wang et al 2017b】

Mnih, V., Heess, N., Graves, A., and Kavukcuoglu, K. (2014). Recurrent models of visual attention. In the Annual Conference on Neural Information Processing Systems

(NIPS)
.

Tips:阅读此文请掌握DQN、Double DQN、Prioritized Experience Replay或多或少二个背景。

Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 380(6266):1332–1338.

Duan, Y., Andrychowicz, M., Stadie, B. C., Ho, J., Schneider, J.,Sutskever, I., Abbeel, P., and Zaremba, W. (2017). One-Shot Imitation Learning. ArXiv e-prints.

Bellemare, M. G., Danihelka, I., Dabney, W., Mohamed, S.,Lakshminarayanan, B., Hoyer, S., and Munos, R. (2017). The Cramer Distance as a Solution to Biased Wasserstein Gradients. ArXiv e-prints.

strategic attentive writer to learn macro-actions 【Vezhnevets et al 2016】

Munos, R., Stepleton, T., Harutyunyan, A., and Bellemare, M. G.(2016). Safe and efficient offpolicy reinforcement learning. In the Annual Conference on Neural Information Processing Systems (NIPS).

Koch, G., Zemel, R., and Salakhutdinov, R. (2015). Siamese neural networks for one-shot image recognition. In the International Conference on Machine Learning (ICML).

Sutton, R. S., Szepesvari, C., and Maei, H. R. (809b). A convergent O( ´ n) algorithm for off-policy temporal-difference learning with linear function approximation. In the Annual Conference on Neural Information Processing Systems (NIPS).

分水岭论文Deep Q-learning Network【Mnih et al 2013】中提到:我真是或多或少人儿的结果看上去很好,或多或少没人 任何理论土最好的办法(原文很狡猾的反过来说一遍)。

Kulkarni, T. D., Narasimhan, K. R., Saeedi, A., and Tenenbaum, J. B. (2016). Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In the Annual Conference on Neural Information Processing Systems (NIPS)

Policy gradient与Q-learning 的结合【O'Donoghue 2017、Nachum 2017、 Gu 2017、Schulman 2017】

Gruslys, A., Gheshlaghi Azar, M., Bellemare, M. G., and Munos, R. (2017). The Reactor: A Sample-Efficient Actor-Critic Architecture. ArXiv e-prints

Barto, A. G., Sutton, R. S., and Anderson, C. W. (1983). Neuronlike elements that can solve difficult learning control problems. IEEE Transactions on Systems, Man, and Cybernetics, 13:835–846

Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., & Riedmiller, M. (2014). Deterministic policy gradient algorithms. International Conference on International Conference on Machine Learning (pp.387-395). JMLR.org.

现有的网络架构搜索土最好的办法【Baker et al 2017、Zoph and Le 2017】,其中Zoph的工作分量非常重。

Lin, L. J. (1993). Reinforcement learning for robots using neural networks.

Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David, Graves, Alex, Antonoglou, Ioannis, Wier- stra, Daan, and Riedmiller, Martin. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5802, 2013.

现有解法围绕着无监督学习开展

Wang, S. I., Liang, P., and Manning, C. D. (2016b). Learning language games through interaction. In the Association for Computational Linguistics annual meeting (ACL)

Tamar, A., Wu, Y., Thomas, G., Levine, S., and Abbeel, P. (2016). Value iteration networks. In the Annual Conference on Neural Information Processing Systems (NIPS).

下面是CV和NLP方面的哪几个简介:物体检测 【Mnih 2014】、机器翻译 【Bahdanau 2015】、图像标注【Xu 2015】、用Attention代替CNN和RNN【Vaswani 2017】等等。

极其优秀的工作:unsupervised reinforcement and auxiliary learning 【Jaderberg et al 2017】

早在1997年Tsitsiklis就证明了很久Function Approximator采用了神经网络或多或少非线性的黑箱,没人 其收敛性和稳定性是无法保证的。

deep exploration via bootstrapped DQN 【Osband et al 2016)】

现有解法有:

Wang, Z., Schaul, T., Hessel, M., van Hasselt, H., Lanctot, M., and de Freitas, N. (2016c). Dueling network architectures for deep reinforcement learning. In the International

Conference on Machine Learning (ICML)
.

Ng, A. and Russell, S. (800).Algorithms for inverse reinforcement learning. In the International Conference on Machine Learning (ICML).

Hester, T. and Stone, P. (2017). Intrinsically motivated model learning for developing curious robots. Artificial Intelligence, 247:170–86.

He, F. S., Liu, Y., Schwing, A. G., and Peng, J. (2017a). Learning to play in a day: Faster deep reinforcement learning by optimality tightening. In the International Conference on Learning Representations (ICLR)

Kaiser, L., Gomez, A. N., Shazeer, N., Vaswani, A., Parmar, N., Jones, L., and Uszkoreit, J. (2017a). One Model To Learn Them All. ArXiv e-prints.

大名鼎鼎的GANs 【Goodfellow et al 2014】

reward function not available

Silver, D., van Hasselt, H., Hessel, M., Schaul, T., Guez, A., Harley, T., Dulac-Arnold, G., Reichert, D., Rabinowitz, N., Barreto, A., and Degris, T. (2016b). The predictron: End-to-end learning and planning. In NIPS 2016 Deep Reinforcement Learning Workshop.

gigantic search space

Taylor, M. E. and Stone, P. (809). Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10:1633–1685.

Vezhnevets, A. S., Mnih, V., Agapiou, J., Osindero, S., Graves, A., Vinyals, O., and Kavukcuoglu, K. (2016). Strategic attentive writer for learning macro-actions. In the Annual Conference on Neural Information Processing Systems (NIPS).

Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. (2016a). Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489.

imitation learning with GANs 【Ho and Ermon 2016、Stadie et al 2017】 (其TensorFlow实现在imitation)

下面跳出DQN的范畴——

现有解法有:

Florensa, C., Duan, Y., and Abbeel, P. (2017). Stochastic neural networks for hierarchical reinforcement learning. In the International Conference on Learning Representations (ICLR)

Sutton, R. S. (1988). Learning to predict by the methods of temporal differences. Machine Learning,3(1):9–44.

GTD 【Sutton 809a、Sutton 809b、Mahmood 2014】

Zoph, B. and Le, Q. V. (2017). Neural architecture search with reinforcement learning. In the International Conference on Learning Representations (ICLR)

Houthooft, R., Chen, X., Duan, Y., Schulman, J., Turck, F. D., and Abbeel, P. (2016). Vime: Variational information maximizing exploration. In the Annual Conference on Neural Information Processing Systems (NIPS).

return-based off-policy control, Retrace 【Munos et al 2016】, Reactor 【Gruslyset al 2017】

Johnson, M., Schuster, M., Le, Q. V., Krikun, M., Wu, Y., Chen, Z., Thorat, N., Viegas, F., Watten- ´berg, M., Corrado, G., Hughes, M., and Dean, J. (2016). Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation. ArXive-prints.

Stadie, B. C., Abbeel, P., and Sutskever, I. (2017).Third person imitation learning. In the International Conference on Learning Representations (ICLR).

旷世猛将van Hasselt先生很喜欢防止Over-Estimate大问题,他先搞出一二个Double Q-learning【van Hasselt 2010】大闹NIPS,六年后搞出强度学习版本的Double DQN【van Hasselt 2016a】!

Q-Prop, policy gradient with off-policy critic 【Gu et al 2017】

model-free planning

DQN的改良主要依靠一二个Trick:

Weiss, K., Khoshgoftaar, T. M., and Wang, D. (2016). A survey of transfer learning. Journal of Big Data, 3(9)

Sutton, R. S., Maei, H. R., Precup, D., Bhatnagar, S., Silver, D., Szepesvari, C., and Wiewiora, ´E. (809a). Fast gradient-descent methods for temporal-difference learning with linear function approximation. In the International Conference on Machine Learning (ICML).

Ravi, S. and Larochelle, H. (2017). Optimization as a model for few-shot learning. In the International Conference on Learning Representations (ICLR).

learn a flexible RNN model to handle a family of RL tasks 【Duan et al 2017、Wang et al 2016a】

focus on salient parts

Gupta, A., Devin, C., Liu, Y., Abbeel, P., and Levine, S. (2017). Learning invariant feature spaces to transfer skills with reinforcement learning. In the International Conference on Learning Representations (ICLR).

比较新的解法有一二个:

新的架构有【Kaiser et al 2017a、Silver et al 2016b、Tamar et al 2016、Vaswani et al 2017、Wang et al 2016c】

TRPO(Trust Region Policy Optimization)【Schulman 2015】

Sutton, R. S., McAllester, D., Singh, S., and Mansour, Y. (800). Policy gradient methods for reinforcement learning with function approximation. In the Annual Conference on Neural Information Processing Systems

(NIPS)
.

data/sample efficiency

或多或少,一二个很好的思路是从计算机视觉与自然语言防止领域汲取灵感,类似 下文中很久提到的unsupervised auxiliary learning土最好的办法借鉴了RNN+LSTM中的絮状操作

benefit from both labelled and unlabelled data

Schulman, J., Levine, S., Moritz, P., Jordan, M. I., and Abbeel, P. (2015). Trust region policy optimization. In the International Conference on Machine Learning (ICML).

model-based learning