Finite-Time Analysis of Whittle Index based Q-Learning for Restless Multi-Armed Bandits with Neural Network Function Approximation

Whittle index policy is a heuristic to the intractable restless multi-armed bandits (RMAB) problem. Although it is provably asymptotically optimal, finding Whittle indices remains difficult. In this paper, we present Neural-Q-Whittle, a Whittle index based Q-learning algorithm for RMAB with neural network function approximation, which is an example of nonlinear two-timescale stochastic approximation with Q-function values updated on a faster timescale and Whittle indices on a slower timescale. Despite the empirical success of deep Q-learning, the non-asymptotic convergence rate of Neural-Q-Whittle, which couples neural networks with two-timescale Q-learning largely remains unclear. This paper provides a finite-time analysis of Neural-Q-Whittle, where data are generated from a Markov chain, and Q-function is approximated by a ReLU neural network. Our analysis leverages a Lyapunov drift approach to capture the evolution of two coupled parameters, and the nonlinearity in value function approximation further requires us to characterize the approximation error. Combing these provide Neural-Q-Whittle with $\mathcal{O}(1/k^{2/3})$ convergence rate, where $k$ is the number of iterations.

Paper

Similar papers

Peer review

Reviewer PM6q7/10 · confidence 4/52023-07-02

Summary

This paper proposes a neural network approach to learning the Whittle index policy. In addition, the paper gives a finite-time analysis for the algorithm, and shows that the algorithm indeed learns the Whittle index values for restless bandits.

Strengths

The paper’s sections are well-written and the technical claims of the paper are sound. In addition, the paper provides sufficient analysis of the two-timescale stochastic approximation used here (to learn the Whittle index values and the action-value function for each arm). Also, the paper offers finite time performance bounds in the neural network setting which is a largely unstudied area.

Weaknesses

While the paper is well-written, I believe it’s better if less remarks/lemmas blocks are used and if they are written as paragraphs. Other than that, the technical details appear correct to the best of my knowledge.

Questions

Question: the neural network is using ReLU activation functions. Would the finite-time analysis be applicable to other activation functions such as tanh?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

The limitations are addressed by the authors in the paper in regards to the finite-time analysis.

Reviewer ZyEQ6/10 · confidence 4/52023-07-02

Summary

This paper presents a neural Q learning method to compute the Whittle indices in restless multi-armed bandit problems. The paper provides an algorithm using two-timescale stochastic approximation (2TSA) to update the parameters in the neural networks and the Whittle indices jointly with different learning rates. The authors also show that the 2TSA method guarantees convergence to the optimal/approximately optimal solution of Whittle indices. One of the major contributions is the breakthrough of not using projection step in the 2TSA algorithm. The algorithm proposed by the authors doesn’t require the functional class to contain the optimal functional approximator. The corresponding approximation guarantee also shows an additional term dependent on the distance of the true optimal to the span of the functional bases used in neural networks. Lastly, the authors provide experiments to show the convergence of the proposed algorithm and empirically verify the assumptions made in the paper.

Strengths

I like the idea of projecting the neural networks using ReLU to linear functions to enable downstream analysis. The paper also generalizes the theoretical analysis in previous work to quantify the impact of approximation error in the functional approximator class. I didn’t go through the appendix, but the proof sketch is clear to me. Overall, the paper is nicely written with new theoretical contribution compared to the previous work. It would be great if the authors can further emphasize and summarize the contributions in the theoretical analysis that are incurred due to not using projection steps. The convergence rate also matches the previous work using projection steps.

Weaknesses

The comparison to the previous algorithm is not clear. I didn’t fully understand why the previous algorithm requires a project step to force the parameters in a bounded set and why your algorithm doesn’t. I believe it is due to the analysis in the convergence guarantee where unbounded parameters can lead to useless bound or divergence. Could you please clarify and emphasis this in the paper more to highlight your contribution more clearly? Especially this seems to be the major contribution of the paper. It deserves a larger portion of the paper to clarify it. Please also clarify why the definition of Whittle index is different from Whittle et al. and most of the literature. Please see below for more details. I am worried that the different definition can impact the convergence analysis (especially the linearity and Lipschitzness in the proof). Please either justify the use of your definition and any references showing this different definition is valid, or show us that the analysis is not impacted by the definition. These two are my major concerns of the paper. Please clarify them and I am happy to update my score based on the response. [Answered by the authors during the rebuttal]

Questions

**Question 1**: Equation 3: the Whittle index defined by Whittle is defined as the Lagrangian multiplier such that the Q values of action 0 and 1 are identical, i.e., whittle index $\coloneqq \inf_\lambda \{ Q_n(s,1;\lambda) = Q_n(s,0;\lambda) \}$, where the Q functions are functions of lambda (Lagrangian multiplier) and thus parameterized by lambda. However, your definition in Equation 3 is defined as the difference between Q values, i.e., whittle index $\coloneqq Q_n(s,0) – Q_n(s,1)$, where the Lagrangian multiplier for defining the Q values is also not specified. To my knowledge, this is not equivalent to the original Whittle index definition. The original Whittle index is defined based on the solution to Lagrangian relaxation, but yours doesn’t have this property. Please clarify/justify why you can define Whittle index in this way or provide any references supporting your claim. And please also clarify how this definition impacts you downstream analysis. I have also checked your references in [3,23] are also aligned with the original Whittle index definition (Whittle index defined such that Q(s,0) = Q(s,1)). Please see Equation (4) in [23] and Equation (11) in [3]. Equation (13) will also be impacted by the definition of Whittle index. So does the analysis in Equation (21) and Lemma 2, which will no longer be linear. References: [3] Konstantin E Avrachenkov and Vivek S Borkar. Whittle index based q-learning for restless bandits with average reward. Automatica, 139:110186, 2022. [23] Jing Fu, Yoni Nazarathy, Sarat Moka, and Peter G Taylor. Towards q-learning the whittle index for restless bandits. In 2019 Australian & New Zealand Control Conference (ANZCC), pages 249–254. IEEE, 2019. --- **Question 2**: Do other choices of step size sequence $\alpha$ and $\eta$ work as well? **Feedback 3**: Equation 18 is the linearized version of Equation 6 where the coefficient associated to the linear term is defined by the initial point $\theta_0 = [w_{r,0}]$. The definition is correct but slightly unclear.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

- Different definition of Whittle index: if this is true, then the analysis might be restricted to the definition shown in the paper. - Linearization: the authors need to linearize the Lyapunov function to enable the theoretical analysis. I believe this is the reason why the authors focus on 2-layer neural networks (only input and output and a ReLU activation function). This can limit the functional approximator class to such neural networks only. For more complex NN structures, the same linearization trick and analysis don't apply anymore.

Reviewer rZoR6/10 · confidence 3/52023-07-07

Summary

This paper investigates the finite-time analysis of the Whittle index-based Q-learning policy for the RMAB problem under neural function approximation. The authors formulate the algorithm as a nonlinear two-time-scale stochastic approximation problem and present a convergence rate of $K^{2/3}$.

Strengths

1. The Neural-Q-Whittle algorithm eliminates the projection step. 2. The paper provides finite-time analysis. 3. Simulation results are included to verify the convergence performance.

Weaknesses

1. It is unclear whether the approximated Q-functions converge or not. 2. The errors diminish as $m\rightarrow \infty$, indicating the need for overparameterization. 3. The simulation setting is too simple to sufficiently demonstrate the advantages of neural approximation. Consideration of a larger state and action space is warranted.

Questions

Since two-time-scale stochastic approximation has been extensively studied recently, could the authors clarify the novel main technical contributions of this paper?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

More limitations of the approach should be added.

Reviewer KAmk7/10 · confidence 3/52023-07-18

Summary

This paper studies Whittle index-based Q-learning with neural network function approximations restless multi-armed bandits (RMAB) problem, which is a model-free low-complexity reinforcement learning (RL) heuristics for RMABs. Since state-action space of RMABs is exponentially growing with the number of arms, common Q-learning techniques as well as tabular Q-Whittle algorithm suffer from curse of dimensionality. Instead, the authors analyze a low-complexity neural network approximation method for Whittle index-based Q-learning under non-iid Markovian state-action observations, namely Neural-Q-Whittle. The authors formulate Neural-Q-Whittle as a nonlinear two-timescale stochastic approximation (2TSA) where the parameters of the neural network Q-function and the Whittle indices are mutually coupled and the former is updated on a faster timescale than the latter. Leveraging a Lyapunov function method, the authors provide a finite-time convergence analysis of Neural-Q-Whittle with non-iid Markovian data. The analysis involves characterizing the error between two Lyapunov functions: one for the neural network Q-function and one for the linear approximation of the neural network.

Strengths

This paper is original in the sense that it removes some of the limitations in the previous theoretical works on finite-time analysis of Whittle index, Q-learning, and nonlinear 2TSA and generalizes to neural network Q-function approximation in two-timescales under non-iid ergodic state-action process. The theoretical analysis and proofs seem technically solid and mostly clear.

Weaknesses

Since this work builds upon the limitations in the related past works, it would be easier to follow and appreciate the novelty and significance if you could provide a table comparing to related works in various aspects (eg. methods, assumptions, convergence rates, computational (time or space) and sample complexities). It could also help with presentation if you give an informal version of your main theorem beforehand or simplify Theorem 1 using asymptotic notation and state the full version in the appendix. I believe it should also be emphasized more that your convergence result relies on diminishing error bound of linear approximation of overparameterized neural networks. Although this is mainly a theory paper, it would be interesting to see the experimental comparison to other methods such as Q-Whittle-LFA and references 13, 22, 58.

Questions

1)I am not sure if I understand the sentence starting with "(Step 3)" between lines 204-207. Can you clarify what you mean exactly? 2) You might include the definition of "span" operator in equation 17 as a footnote. 3) Are h, g, y in Remark 4 supposed be h_0, g_0, y_0? 4) It is not exactly clear to me why it is desirable to update neural network parameters and the indices at two different time-scales. Do these parameters only converge under this case? Or do they converge faster compared to single timescale? Could you clarify the need for and the significance of two-timescale approximation?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

The limitations are addressed within the text and mentioned as a future work in the conclusion (extension to multi-layer neural network).

Reviewer bjcG6/10 · confidence 3/52023-07-26

Summary

This paper provides a finite-time analysis of Neural-Q-Whittle. The authors propose Neural-Q-Whittle, a novel Whittle index-based Q-learning algorithm with neural network function approximation for RMAB. Their analysis leverages a Lyapunov drift approach to capture the evolution of two coupled parameters, and the nonlinearity in value function approximation further requires us to characterize the approximation error. They also conduct experiments to validate the convergence performance of Neural-Q-Whittle, and verify the sufficiency of their proposed condition for the stability of Neural-Q-Whittle.

Strengths

1. This paper provides a non-asymptotic convergence rate analysis of Neural-Q-Whittle with two coupled parameters updated in two timescales under Markovian observations without the extra projection step. 2. The authors propose Neural-Q-Whittle, a novel Whittle index-based Q-learning algorithm with neural network function approximation for RMAB. 3. The authors establish the first finite-time analysis of Neural-Q-Whittle under Markovian observations. Their analysis leverages a Lyapunov drift approach to capture the evolution of two coupled parameters, and the nonlinearity in value function approximation further requires us to characterize the approximation error. 4. The writing is good.

Weaknesses

Compared with existing works [13, 22, 58] for Q-learning with neural network function approximation, the technique contribution of this paper is kind of limited. It will be better if the authors could include additional parts to clarify the technical contribution of the paper, compared with prior works.

Questions

Please refer to the Weaknesses section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No

Reviewer ZyEQ2023-08-13

Thank you for your response

I appreciate the authors for their detailed response, especially the response to my Question #2. It addresses my concern about the definition of Whittle index and I will update the rating. Thank you!

Authorsrebuttal2023-08-13

Thank you!

Thank you for your acknowledgement and raising the rating of our paper. Much appreciated!

Reviewer PM6q2023-08-15

I thank the authors for answering my question; I will keep my score at 7 for the paper.

Authorsrebuttal2023-08-15

Thank you!

Thank you for your acknowledgement and keeping the positive rating of our paper. Much appreciated!

Reviewer rZoR2023-08-16

Thanks for your response and clarifying some of my questions. I would like to increase the rating.

Authorsrebuttal2023-08-16

Thank you!

Thank you for your acknowledgement and raising the rating of our paper. Much appreciated!

Authorsrebuttal2023-08-17

Thank you!

Thank you for your acknowledgement and keeping the positive rating of our paper. Much appreciated!

Reviewer KAmk2023-08-22

Thank you for taking time and effort in addressing my questions and concerns and clarifying some of the confusing points. I am satisfied with your response and will keep my score.

Authorsrebuttal2023-08-22

Thank you!

Thank you for your acknowledgement and keeping the positive rating of our paper. Much appreciated!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC