Weakly Coupled Deep Q-Networks

We propose weakly coupled deep Q-networks (WCDQN), a novel deep reinforcement learning algorithm that enhances performance in a class of structured problems called weakly coupled Markov decision processes (WCMDP). WCMDPs consist of multiple independent subproblems connected by an action space constraint, which is a structural property that frequently emerges in practice. Despite this appealing structure, WCMDPs quickly become intractable as the number of subproblems grows. WCDQN employs a single network to train multiple DQN "subagents", one for each subproblem, and then combine their solutions to establish an upper bound on the optimal action value. This guides the main DQN agent towards optimality. We show that the tabular version, weakly coupled Q-learning (WCQL), converges almost surely to the optimal action value. Numerical experiments show faster convergence compared to DQN and related techniques in settings with as many as 10 subproblems, $3^{10}$ total actions, and a continuous state space.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer QmyG7/10 · confidence 4/52023-07-08

Summary

The paper studies value-based control agents for Weakly Coupled MDPs (WCMDPs), which comprise independent subproblems linked by an action constraint. The authors generalize an existing Lagrangian relaxation approach to action-value methods learned by stochastic approximation to obtain Weakly Coupled Q-Learning (WCQL). Essentially, a Lagrange multiplier is used to create an overestimate of $Q^*$ called $Q^{\lambda*}$; taking the minimum of these two value functions can lead to faster convergence without changing the fixed point. A deep-learning extension of this method relying on a soft constraint added to the DQN loss yields the titular method Weakly Coupled Deep Q-Network (WCDQN). The proposed methods are tested against several well-known baselines (such as DQN and Double DQN) and compare favorably on three WCMDPs: electric vehicle charging (tabular only), multi-product inventory control, and ad matching problems.

Strengths

- Self-contained descriptions of Weakly Coupled MDPs and Lagrangian relaxation, as well as a complete related works discussion, make the paper accessible to people (like me) who do not normally work in this setting. The explanations are clear and well written. - Interesting theoretical results in the tabular setting. The convergence results look correct and intuitive to me. I checked the proofs in the appendix. - Interesting, realistic experiment domains and strong empirical results.

Weaknesses

- WCQL relies on discretizing $\lambda \geq 0$ into a finite set of values $\Lambda$ to approximate the upper bound $\min_{\lambda} Q^{\lambda}$. Although this does not affect the ultimate convergence to $Q^*$ in the tabular case, I would imagine that this choice has a significant impact on empirical performance, yet the authors do not discuss this or how to choose the set (that I was able to find at least). - The computational requirements of the proposed methods are downplayed. The authors state, “We note that, especially in settings where the limiting factor is the ability to collect enough experience, these bounds come nearly for free in that they only require computing a few additional quantities and importantly, they do not require additional experience.” I think this is a strong claim; compared to Q-Learning, at least $N$ additional Q-functions are needed to form the upper bound $Q^{\lambda*}$. In the deep RL setting, only one extra network appears to be needed, in order to generalize over these $N$ functions, but that still doubles the overall computation. Thus, I do not think it is fair to say that the method comes for free; however, the strong empirical results do justify the extra expense. - Although the tabular theory is sound, the implications of porting it to deep function approximation are unclear and not fully appreciated by the paper in my opinion. The authors say, “the [loss] penalty softly ensures that learned network satisfies the bounds obtained from the Lagrangian relaxation.” However, because the constraint is soft, it does not ensure anything. If the approximated upper bound ever dips below Q*, then the wrong value will be bootstrapped by the network and the learned returns will be suboptimal. Of course, this is inevitable with function approximation (a hard constraint cannot be enforced), but it is worthwhile to clarify in the paper.

Questions

- Why is $\lambda$ called the penalty vector? Is it not a scalar, i.e., the Lagrange multiplier? - How long (in timesteps) were the agents trained in each of the experiments? Since the figure x-axes use episodes, did agents receive unequal training time? - How was the penalty coefficient $\kappa_U=10$ chosen?

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

- Need to choose the set $\Lambda$ - Added computational expense - Upper bound potentially violated with function approximation and soft constraint

Reviewer da8o6/10 · confidence 3/52023-07-10

Summary

The authors present a new algorithm, "WCDQN" for solving structured problems know as weakly coupled MDPs. The method dynamically estimates upper bounds via Lagrangian relaxation essentially to improve the value of collected experience across the structured subproblems. The authors present a (re)formulation of the problem via the dual bound of the linking constraints and numerically demonstrate the effectiveness of the method on a set of WCMDP problems.

Strengths

1. Nicely written, enjoyable reading 2. Relatively simple and clear problem formulation

Weaknesses

1. moderate ablation. the proposed method allows for a decomposition based on the number of "subproblems" but the relative effect to this key variable is not explicitly tested

Questions

1. It would seem to be the case that this method should scale significantly better for problems with specific substructure. Why not demonstrate this explicitly with a custom problem / domain to validate the extreme case? 2. [note] I think Fig3 is missing a y-axis? 3. [note] your compilation resulted in non-clickable links,

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

NA - nothing in addition to what is present in the actual mathematics.

Reviewer HJ3Y5/10 · confidence 3/52023-07-10

Summary

This paper considers the problem of learning optimal policies for a weakly coupled Markov decision process (WCMDP). The paper extends the classical DQN algorithm to this setting by introducing separate lagrangian relaxations. Experiments are conducted on 3 operational research testbeds.

Strengths

This paper has a very complete structure, with a detailed problem introduction, an insightful algorithm derivation, a theoretical analysis as well as solid numerical studies. The presentation is particularly clear and easy to follow. The algorithm is intuitive and convincing. You cannot really find any flow from this paper.

Weaknesses

My biggest concern about this work is the audience. As a practitioner in deep RL, this is literally the first time for me to hear the problem of WCMDP. Although I do appreciate that this work teaches me a lot, I don't know whether this paper can attract the attention of the deep RL community, particularly considering the test domains are pretty toy. This paper looks to me more like a work that brings deep RL techniques to the OR community rather than the reverse direction.

Questions

N/A

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

4 excellent

Contribution

3 good

Limitations

The significance of the problem for the deep learning community.

Reviewer aL5V5/10 · confidence 3/52023-07-12

Summary

This paper introduces novel Q-learning based approaches (WCQL and WCDQN) to solving weakly-coupled Markov decision processes. The approaches work by running Q-learning for the sub-problems of the WCMDP to construct an upper bound on the optimal Q-values of the overall problem. The WCMDP constraint is incorporated into the reward structure of the task via approximate Lagrangian relaxation. In the function approximation case, the constraint is enforced as a penalty to the loss. A theoretical analysis is performed for the tabular case and experiments are run to demonstrate the performance of both the tabular and DQN-based approaches.

Strengths

- The paper is generally easy to follow and clearly written, and the problem setting is well-motivated. - The proposed approaches appear to be simple and relatively original. - If this is indeed the first model-free approach to WCMDPs, it represents a significant advance for this problem setting.

Weaknesses

- The core weakness of both the theoretical and empirical analyses, in my opinion, is the lack of treatment of constraint violations by WCQL and WCDQN. The theoretical analyses, as far as I can tell, assume a fixed, arbitrary Lagrange multiplier and show convergence of the Q-values to the optimal Q-values for *that choice of Lagrange multiplier*. This, however, says nothing about the resulting policy’s constraint violations. Similarly, in the experiments, there are no results showing the constraint violations of each method. Indeed, the “soft constraint” formulation seems strange for problems like the EV task, for in which in practice assigning too many cars to too few charging ports would seem to be a serious problem. I think in particular it would be helpful to add plots showing constraint violations for each method and whether they find feasible policies. - Related to the above, there’s no discussion of whether even in an idealized case where $\lambda$ was a continuous variable the proposed approach converges to a saddle point of the Lagrangian (i.e., actually solves the “true” optimization problem), even if in practice discretization is not a significant detriment to performance. - As far as I can see, there is no discussion of the limitations of the proposed approaches. - The experimental evaluation is fairly limited. - More details on the baselines in the main text would be helpful to improve clarity. I would suggest changing the citations to a numerical format to save space. Minor: - The stochastic gradient update at the end of Section 3.2 would ideally reflect the update of minibatch SGD rather than pure SGD. - The $\wedge$ symbol in Eq. 11 seems unnecessary, it would be more clear to just write $\min(\cdot, \cdot)$. - From a clarity perspective, it would be helpful if the initial description of a WCMDP contained an example to anchor readers’ understanding. - Figure 3A has no y-axis label. Also, I think it would be more informative to plot the upper bound on average across $(s,a)$ pairs rather than at only one point. - It would make sense to add distributional RL [1] and Rainbow [2] to the list of extensions of DQN in the introduction or related work sections. The authors should also contrast their setting to constrained MDPs [3]. I think this is interesting work, and I am open to increasing my score based on the degree to which the above concerns are addressed. [1] https://arxiv.org/abs/1707.06887 [2] https://arxiv.org/abs/1710.02298 [3] https://www-sop.inria.fr/members/Eitan.Altman/TEMP/h.pdf

Questions

- Does the sub-problem action-value network parameterization mean that all of the subproblem action spaces have to have the same number of actions? - Is there some intuition as to why double Q-learning performs worse than regular Q-learning? I would think the fact that double Q-learning attempts to reduce value overestimation would have a similar effect to ensuring the overall task Q-value estimate doesn’t exceed the estimated upper bound.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Contribution

3 good

Limitations

The authors do not appear to discuss any limitations of their method (see above). I do not directly see any negative societal implications of this work, though in practice it would depend on the problem to which it is applied and any guarantees (or lack thereof) regarding constraint violations.

Reviewer HJ3Y2023-08-12

Thanks for the clarifications. I will keep a positive perspective on this work and leave the final decision to the discussion phase.

Authorsrebuttal2023-08-13

Thank you for reading our rebuttal and for the kind comments! Please let us know if you have any further questions about the applicability of the weakly-coupled technique to real-world problems that the deep RL community is interested in. We do believe that we are both bringing an OR technique (the weakly-coupled formulation + the decomposition and upper bound) to the deep RL community, while also bringing deep RL ideas (DQN agents) to the OR community and hope that this paper can be beneficial to both sides.

Reviewer aL5V2023-08-12

Response

Thank you to the authors for your detailed rebuttal! It has clarified a number of my questions. I think the added discussion should will benefit the text as well. I'm happy to raise my score.

Authorsrebuttal2023-08-13

Thank you for the acknowledging that we have helped to clarify some questions and for raising the score! We appreciate your time. Please let us know if we can further expand or clarify any points that we were not able to fully discuss in the rebuttal. We'd love to have the opportunity to address any remaining concerns you have in the next few days before the discussion period ends.

Reviewer QmyG2023-08-16

Thank you for the detailed clarifications! Adding these points of discussion to the paper will help others understand your algorithm design and experiment choices better. I will keep my score the same and continue to recommend this paper for acceptance.

Authorsrebuttal2023-08-16

We appreciate the response and your time spent reviewing the paper! Please let us know if any additional questions arise.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC