Abstract
We consider model selection for classic Reinforcement Learning (RL) environments – Multi Armed Bandits (MABs) and Markov Decision Processes (MDPs) – under general function approximations. In the model selection framework, we do not know the function classes, denoted by \(\mathcal {F}\) and \(\mathcal {M}\), where the true models – reward generating function for MABs and transition kernel for MDPs – lie, respectively. Instead, we are given M nested function (hypothesis) classes such that true models are contained in at-least one such class. In this paper, we propose and analyze efficient model selection algorithms for MABs and MDPs, that adapt to the smallest function class (among the nested M classes) containing the true underlying model. Under a separability assumption on the nested hypothesis classes, we show that the cumulative regret of our adaptive algorithms match to that of an oracle which knows the correct function classes (i.e., \(\mathcal {F}\) and \(\mathcal {M}\)) a priori. Furthermore, for both the settings, we show that the cost of model selection is an additive term in the regret having weak (logarithmic) dependence on the learning horizon T.
A. Ghosh and S. R. Chowdhury contributed equally.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Similar content being viewed by others
Notes
- 1.
Here the roles of \(x_1\) and \(x_2\) are interchangeable without loss of generality.
- 2.
We assume that the action set \(\mathcal {X}\) is compact and continuous, and so such action pairs \((x_1,x_2)\) always exist, i.e., given any \(x_1 \in \mathcal {X}\), an action \(x_2\) such that \(D^*(x_1,x_2) \le \eta \) always exists.
- 3.
This can be found using standard trick like doubling.
- 4.
For any \(\alpha > 0\), we call \(\mathcal {F}^\alpha \) an \((\alpha ,\left\Vert \cdot \right\Vert _{\infty })\) cover of the function class \(\mathcal {F}\) if for any \(f \in \mathcal {F}\) there exists an \(f'\) in \(\mathcal {F}^\alpha \) such that \(\left\Vert f' - f\right\Vert _{\infty }:=\sup _{x \in \mathcal {X}}|f'(x)-f(x)|\le \alpha \).
- 5.
We can extent the range to [0, c] without loss of generality.
- 6.
One can choose \(\delta = 1/\text {poly}(M)\) to obtain a high-probability bound which only adds an extra \(\log M\) factor.
- 7.
For any \(\alpha > 0\), \(\mathcal {P}^\alpha \) is an \((\alpha ,\left\Vert \cdot \right\Vert _{\infty ,1})\) cover of \(\mathcal {P}\) if for any \(P \in \mathcal {P}\) there exists an \(P'\) in \(\mathcal {P}^\alpha \) such that \(\left\Vert P' - P\right\Vert _{\infty ,1}:=\sup _{s,a}\int _{\mathcal {S}}|P'(s'|s,a)-P(s'|s,a)|ds' \le \alpha \).
References
Abbasi-Yadkori, Y., Pál, D., Szepesvári, C.: Improved algorithms for linear stochastic bandits. In: Advances in Neural Information Processing Systems, pp. 2312–2320 (2011)
Agarwal, A., Luo, H., Neyshabur, B., Schapire, R.E.: Corralling a band of bandit algorithms. In: Conference on Learning Theory, pp. 12–38. PMLR (2017)
Arora, R., Marinov, T.V., Mohri, M.: Corralling stochastic bandit algorithms. In: International Conference on Artificial Intelligence and Statistics, pp. 2116–2124. PMLR (2021)
Ayoub, A., Jia, Z., Szepesvari, C., Wang, M., Yang, L.F.: Model-based reinforcement learning with value-targeted regression. arXiv preprint arXiv:2006.01107 (2020)
Balakrishnan, S., Wainwright, M.J., Yu, B., et al.: Statistical guarantees for the EM algorithm: from population to sample-based analysis. Ann. Stat. 45(1), 77–120 (2017)
Besbes, O., Zeevi, A.: Dynamic pricing without knowing the demand function: risk bounds and near-optimal algorithms. Oper. Res. 57(6), 1407–1420 (2009)
Cesa-Bianchi, N., Lugosi, G.: Prediction, Learning, and Games. Cambridge University Press, Cambridge (2006)
Chang, F., Lai, T.L.: Optimal stopping and dynamic allocation. Adv. Appl. Probabil. 19(4), 829–853 (1987). http://www.jstor.org/stable/1427104
Chatterji, N.S., Muthukumar, V., Bartlett, P.L.: Osom: a simultaneously optimal algorithm for multi-armed and linear contextual bandits. arXiv preprint arXiv:1905.10040 (2019)
Chowdhury, S.R., Gopalan, A.: Online learning in kernelized Markov decision processes. In: The 22nd International Conference on Artificial Intelligence and Statistics, pp. 3197–3205 (2019)
Even-Dar, E., Mannor, S., Mansour, Y.: Action elimination and stopping conditions for the multi-armed bandit and reinforcement learning problems. J. Mach. Learn. Res. 7(39), 1079–1105 (2006). http://jmlr.org/papers/v7/evendar06a.html
Foster, D.J., Krishnamurthy, A., Luo, H.: Model selection for contextual bandits. In: Advances in Neural Information Processing Systems, pp. 14714–14725 (2019)
Ghosh, A., Chowdhury, S.R., Gopalan, A.: Misspecified linear bandits. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 31 (2017)
Ghosh, A., Pananjady, A., Guntuboyina, A., Ramchandran, K.: Max-affine regression: Provable, tractable, and near-optimal statistical estimation. arXiv preprint arXiv:1906.09255 (2019)
Ghosh, A., Sankararaman, A., Kannan, R.: Problem-complexity adaptive model selection for stochastic linear bandits. In: International Conference on Artificial Intelligence and Statistics, pp. 1396–1404. PMLR (2021)
Ghosh, A., Sankararaman, A., Ramchandran, K.: Model selection for generic contextual bandits. arXiv preprint arXiv:2107.03455 (2021)
Jin, C., Yang, Z., Wang, Z., Jordan, M.I.: Provably efficient reinforcement learning with linear function approximation. arXiv preprint arXiv:1907.05388 (2019)
Kakade, S., Krishnamurthy, A., Lowrey, K., Ohnishi, M., Sun, W.: Information theoretic regret bounds for online nonlinear control. arXiv preprint arXiv:2006.12466 (2020)
Krishnamurthy, S.K., Athey, S.: Optimal model selection in contextual bandits with many classes via offline oracles. arXiv preprint arXiv:2106.06483 (2021)
Lee, J.N., Pacchiano, A., Muthukumar, V., Kong, W., Brunskill, E.: Online model selection for reinforcement learning with function approximation. CoRR abs/2011.09750 (2020). https://arxiv.org/abs/2011.09750
Mnih, V., et al.: Playing Atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)
Osband, I., Van Roy, B.: Model-based reinforcement learning and the eluder dimension. In: Advances in Neural Information Processing Systems 27 (NIPS), pp. 1466–1474 (2014)
Pacchiano, A., Dann, C., Gentile, C., Bartlett, P.: Regret bound balancing and elimination for model selection in bandits and RL. arXiv preprint arXiv:2012.13045 (2020)
Pacchiano, A., et al.: Model selection in contextual stochastic bandit problems. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems, vol. 33, pp. 10328–10337. Curran Associates, Inc. (2020). https://proceedings.neurips.cc/paper/2020/file/751d51528afe5e6f7fe95dece4ed32ba-Paper.pdf
Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley, New York (2014)
Russo, D., Van Roy, B.: Eluder dimension and the sample complexity of optimistic exploration. In: Advances in Neural Information Processing Systems, pp. 2256–2264 (2013)
Silver, D., et al.: Mastering the game of go without human knowledge. Nature 550(7676), 354–359 (2017)
Srinivas, N., Krause, A., Kakade, S.M., Seeger, M.W.: Information-theoretic regret bounds for gaussian process optimization in the bandit setting. IEEE Trans. Inf. Theory 58(5), 3250–3265 (2012)
Wang, R., Salakhutdinov, R., Yang, L.F.: Provably efficient reinforcement learning with general value function approximation. arXiv preprint arXiv:2005.10804 (2020)
Wang, Y., Wang, R., Du, S.S., Krishnamurthy, A.: Optimism in reinforcement learning with generalized linear function approximation. arXiv preprint arXiv:1912.04136 (2019)
Williams, G., Aldrich, A., Theodorou, E.A.: Model predictive path integral control: From theory to parallel computation. J. Guid. Control. Dyn. 40(2), 344–357 (2017)
Yang, L.F., Wang, M.: Reinforcement leaning in feature space: matrix bandit, kernels, and regret bound. arXiv preprint arXiv:1905.10389 (2019)
Yang, Z., Jin, C., Wang, Z., Wang, M., Jordan, M.: Provably efficient reinforcement learning with kernel and neural function approximations. In: Advances in Neural Information Processing Systems, vol. 33 (2020)
Yi, X., Caramanis, C., Sanghavi, S.: Solving a mixture of many random linear equations by tensor decomposition and alternating minimization. CoRR abs/1608.05749 (2016). http://arxiv.org/abs/1608.05749
Zanette, A., Brandfonbrener, D., Brunskill, E., Pirotta, M., Lazaric, A.: Frequentist regret bounds for randomized least-squares value iteration. In: International Conference on Artificial Intelligence and Statistics, pp. 1954–1964 (2020)
Acknowledgements
We thank anonymous reviewers for their useful comments. Moreover, we would like to thank Prof. Kannan Ramchandran (EECS, UC Berkeley) for insightful discussions regarding the topic of model selection. SRC is grateful to a CISE postdoctoral fellowship of Boston University.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
1 Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Copyright information
© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Ghosh, A., Chowdhury, S.R. (2023). Model Selection in Reinforcement Learning with General Function Approximations. In: Amini, MR., Canu, S., Fischer, A., Guns, T., Kralj Novak, P., Tsoumakas, G. (eds) Machine Learning and Knowledge Discovery in Databases. ECML PKDD 2022. Lecture Notes in Computer Science(), vol 13716. Springer, Cham. https://doi.org/10.1007/978-3-031-26412-2_10
Download citation
DOI: https://doi.org/10.1007/978-3-031-26412-2_10
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-031-26411-5
Online ISBN: 978-3-031-26412-2
eBook Packages: Computer ScienceComputer Science (R0)