Dual Lagrangian Learning for Conic Optimization

This paper presents Dual Lagrangian Learning (DLL), a principled learning methodology for dual conic optimization proxies. DLL leverages conic duality and the representation power of ML models to provide high-duality, dual-feasible solutions, and therefore valid Lagrangian dual bounds, for linear and nonlinear conic optimization problems. The paper introduces a systematic dual completion procedure, differentiable conic projection layers, and a self-supervised learning framework based on Lagrangian duality. It also provides closed-form dual completion formulae for broad classes of conic problems, which eliminate the need for costly implicit layers. The effectiveness of DLL is demonstrated on linear and nonlinear conic optimization problems. The proposed methodology significantly outperforms a state-of-the-art learning-based method, and achieves 1000x speedups over commercial interior-point solvers with optimality gaps under 0.5\% on average.

Paper

Similar papers

Peer review

Reviewer H5sy7/10 · confidence 3/52024-07-04

Summary

The authors propose Dual Langrangian Learning for learning dual conic optimization problems. Dual conic completion, differentiable conic projection layers, and a self supervised Lagrangian training framework are discussed.

Strengths

- Presents a unique framework for the dual optimization. No dedicated framework for dual optimization proxies are known to date -Extensive theoretical formulations

Weaknesses

- The practical use of the framework is unclear especially with regard to scalability and computational costs involved

Questions

- It is unclear how the utility of this framework can be tested for problem-solving in the real world. While DC3 has been selected as a comparable technique, it remains to be seen whether dual optimization is always beneficial -- for e.g. assuming a problem can be solved both in the primal and dual, are there necessary computational benefits to solving in the dual? How scalable are the proposed algorithmic frameworks?

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

No.

Reviewer H5sy2024-08-13

I thank the authors for their clarifying comments. I choose to keep my original scores for the paper.

Reviewer vMXP7/10 · confidence 4/52024-07-06

Summary

This paper addresses the broad category of machine learning (ML) for optimization, where ML-based approaches are used to obtain solutions or bounds on the solutions of optimization problems. In particular, the paper addresses a particular subclass of problem - conic problems, and even more specifically, the dual solution of these problems. Most "ML-for-optimization" frameworks have thus far focused on finding the solution to primal problems, which is intuitive because the variables in primal problems are typically the variables of interest for most real-world problems. However, learning dual solutions can help certify or bound the primal solution, and in some cases, one may be interested in the dual solution directly (for example, when learning shadow prices in economic optimization problems). This paper uses a conic projection layer in addition to the dual completion step to ensure dual feasibility. The paper compares the method (DLL) to other existing methods in terms of their ability to learn solutions to dual problems and shows superiority, mostly due to the fact that DLL includes dual optimality conditions in addition to dual feasibility constraints. An alternative to Euclidean projection is also provided (Radial projection), for the projection layer, which allows for computation once per cone and attempts to avoid gradient vanishing issues during training.

Strengths

The paper claims that the model can guarantee dual feasibility, which is significant for yielding bounds on the primal solution. In addition, the authors show a closed-form, analytical solution for conic projections and for dual conic completion, outside of the ML model development. The proposed DLL framework doesn't require an explicit formulation of the dual problem; instead, the user must only have to provide a set of primal constraints that satisfy a certain set of conditions. This is convenient but not a main contribution (if the primal conic problem is known, the dual should be relatively easy to find). Even though the most challenging and time consuming optimization problems in practice are nonconvex optimization problems, many of these problems can be cast as conic problems, and thus the method can be applied to solve quite a large array of problems that may occur in practice. Versus other methods, the DLL framework requires far less hyperparameter tuning and iterative procedures, which contributes to the ease of training and increased computational speed.

Weaknesses

One limitation of the method, if a known closed-form optimal dual completion cannot be achieved, is its use of implicit layers. It would be interesting to see how the speedups diminish with very large problems where the implicit layer becomes a bottleneck. The comparison with existing approaches is fair since those are the state of the art for learning primal solutions, but it's slightly unfair because those methods were not specifically designed to learn dual solutions. I think this is fine since there are not existing ML-for-optimization methods tailored towards dual solutions in particular, but just to note. The work in this paper does seem like a more generalized and extended version of the work in reference [19], where reference [19] was the first to provide a dual proxy for a second-order cone relaxation of the AC optimal power flow (OPF) problem. However, in reading [19], it does not seem to be necessarily limited to the AC OPF problem, although those authors only apply it to such. The approach in the present paper is generalized to a broader class of problems, but is not the first to provide dual feasibility guarantees as determined by a learning model.

Questions

It would be interesting to also see indications of whether or not strong duality holds in the test cases, if this is not too challenging to check. If so, the optimality gaps reported also should hold in the primal. The authors may also want to come up with some examples of why learning dual solutions directly may be a contribution on its own, outside of just providing a bound on a primal solution. I would also want to see more of a discussion about the comparison with reference [19], specifically, is [19] already generalizable but the authors only applied it to the AC OPF problem in the paper? Regarding the use of implicit layers, a discussion (or even better, experiments if time allows) on their computational limitations and how this slows down inference as problem size grows

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have sufficiently addressed the limitations of the method. I don't see any potential negative societal impacts of the work.

Reviewer 9KhZ7/10 · confidence 4/52024-07-13

Summary

This paper presents the Dual Lagrangian Learning (DLL) framework that utilizes a fully-connected neural network (FCNN) to provide high-quality dual-feasible solutions that can be used to generate valid Lagrangian dual bounds for linear and nonlinear conic optimization problems without the use of costly implicit layers. This FCNN provides a candidate dual feasible point that is passed through a conic projection step to produce a dual feasible point, then dual completion subproblem is solved to get a complete dual feasible pair $(\hat{z}, \hat{y}) which is then used to get a valid Lagrangian dual bound. The subgradient of the Lagrangian at the dual feasible point is used to update the FCNN network weights to provide better candidate points. The dual completion subproblem is solved for a variety of conic optimization problems. The experimental evaluation is done on both linear and nonlinear conic optimization problems and it is shown to outperform the state-of-art method DC3.

Strengths

The strength of the paper is the simplicity and effectiveness of the DLL framework compared to DC3 and commercial interior point solvers. While previous work used dual feasible points as a warm-start, this work shows how dual feasible points can be used to get valid Lagrangian dual bounds.

Weaknesses

The main weakness of the paper is lack of consideration for general non-convex constraints. A minor weakness of the paper is lack of detail about how DC3 was tuned. The authors claim that DC3 was difficult to tune and only limited tuning was done with DC3 due to this. However, on the positive side, they do provide adequate detail of how the hyperparameters for DC3 were set for easy reproducibility.

Questions

1. On Line 515-6, there seems to be a typo. "do not valid bounds" should be replaced with "do not output valid bounds" 2. How was the limited tuning of DC3 done? While DC3 has many parameters and is difficult to tune, it is possible that DC3 maybe better with more extensive tuning. It is important to show that a reasonable effort was made to tune DC3 before claiming that its performance is worse than DLL. 3. How does the quality of the candidates provided by DLL change over time? It would be interesting to see how the Lagrangian dual bound changes with each iteration of DLL. 4. What considerations did you use when designing the FCNN for DLL? Have you tried other NN architectures other than GNNs?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors acknowledge the limitations of their work, which include lack of consideration for general non-convex constraints and suggest Graph Neural Networks (GNNs) as a promising future avenue of research.

Reviewer yA4S3/10 · confidence 5/52024-07-14

Summary

The paper introduces Dual Lagrangian Learning (DLL), a machine learning framework designed for dual conic optimization problems. DLL utilizes conic duality and machine learning models to produce high-quality, dual-feasible solutions and corresponding Lagrangian dual bounds for both linear and nonlinear conic optimization problems. It introduces techniques like a systematic dual completion procedure, differentiable conic projection layers, and a self-supervised learning framework based on Lagrangian duality. The methodology is demonstrated to significantly outperform existing state-of-the-art methods and provides a faster alternative to traditional interior-point methods used in commercial solvers.

Strengths

The paper introduces Dual Lagrangian Learning (DLL), which combines machine learning (ML) techniques with dual conic optimization to enhance the computation of dual feasible solutions and corresponding Lagrangian bounds. This integration addresses both linear and nonlinear conic optimization problems, aiming at providing a new approach as compared to a traditional methods. The methodology applies differentiable programming to optimization, thus enabling gradient-based optimization methods to refine dual solutions directly. The paper is well-structured and articulates complex ideas with clarity.

Weaknesses

Despite its strengths, the paper has several areas that could be improved: Scope of Application: The focus on continuous problems seems a step back considering the current trend and necessity of addressing mixed-integer programming (MIP) in real-world important applications like power systems and manufacturing, which authors rightfully point out in the beginning. The utility of DLL in these contexts remains unclear, especially without addressing the discrete aspects of such problems. Incremental Nature: The contributions, while innovative in combining ML with dual conic optimization, appear incremental when compared to existing research integrating ML with Lagrangian relaxation for MIP. More distinction or comparison with these methods could help in highlighting the unique advantages of DLL. Numerical Validation: The experiments provided are based on relatively small problem instances. This limitation might raise concerns about the scalability and practical applicability of DLL. Even by the standards of MIP applications, the sizes of the problems tested appear to be rather small. Extending these experiments to larger, more complex problem instances could enhance the paper's impact and relevance.

Questions

Justification for Focus on Continuous Problems: Can the authors provide a stronger rationale for focusing predominantly on continuous problems given the prevalent interest and necessity for solving MIP in real-world applications? Comparison with Existing Methods: How does DLL compare in performance and applicability with existing methods that combine ML with Lagrangian relaxation specifically for MIP problems? Scalability and Practical Applicability: Could the authors elaborate on the potential scalability of DLL? Are there plans to test the methodology on larger-scale problems, particularly those involving discrete decisions common in industrial applications?

Rating

3

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

The limitations are provided in the paper, but in the view of the above review, they appear to be rather restrictive.

Reviewer yA4S2024-08-09

I thank the authors for clarifying their stance. There are the following points that need to be addressed: 1. Presentation-wise, the language needs to be adjusted to make sure that the authors are not overclaiming, For example, a good lower bound can be obtained to certify the solution quality, but the solution itself may not necessarily be obtained. 2. The paper's contribution must be put in the context of the existing research integrating ML with Lagrangian relaxation. 3. The authors consider a subgradient variant of the method, which may suffer from zigzagging. This loops back to point #2 above, necessitating a more in-depth literature review and, likely, more testing, time permitting.

Authorsrebuttal2024-08-12

We thank the reviewer for their additional feedback. We will update the last sentence in the second paragraph of Section I > it becomes possible to design a primal proxy to deliver a high-quality primal solution and an associated dual proxy to obtain a quality certificate. to make it clear that the paper focuses on the dual side, i.e., the primal proxy is trained separately. We would like to point out that the contributions listed in Section 1.1 specifically mention only conic problems, from conic duality to conic projection and conic problems in the experiments. We will include, in Section 6, an additional discussion about the non-convex and discrete settings, and will expand the literature review in Section 2 to mention additional works on Lagrangian relaxation in the MIP context. The dual Lagrangian function is concave (for a minimization problem) and sub-differentiable. Hence, most algorithms for Lagrangian relaxation are based on subgradients, unless a regularizer is used as in, e.g., Ref [20]. The zig-zagging issue is problematic, in an optimization context, when solving a single instance. In the ML context, as is the case here, the dual Lagrangian function is used as training loss function. In particular, it is evaluated (and averaged) across multiple instances in a minibatch, which provides a form or regularization. It is also important to note that 1) popular activation functions like ReLU also yield subgradients, without significant impact on performance, and 2) the models are trained with the Adam optimizer which uses momentum, thus reducing the risk of zigzagging. Finally, we note that works on Lagrangian relaxation in the MIP setting also use subgradient-based training, e.g., [this paper](https://openreview.net/forum?id=aZnZOqUOHq).

Reviewer 9KhZ2024-08-12

I thank the authors for their detailed response to my comments and questions. 1. Regarding the general rebuttal about lack of consideration for non-convex constraints, I agree with the author's comments that convex case is a good starting point as convex relaxations are typically used for non-convex constraints. I recommend that the authors put a non-convex use case if submitting elsewhere as it may strengthen the paper. 2. Regarding the tuning of DC3, I thank the authors for providing a detailed account of the tuning of DC3 in the comments. It is understandable that extensive tuning costs time and computational resources. Still, I would recommend that the authors put these details in the paper to justify the limited tuning of DC3 and explain how the final parameters were chosen in the paper. 3. Regarding the consideration of other architectures other than chosen FCNN architectures, I was more interested in what other architectures the authors experimented with and why this particular architecture was chosen. 4. Good to hear that the authors will consider putting a figure comparing the evolution of the DLL and DC3 bound during training. It will be a useful addition to the paper. With the addressing of the above comments, I would like to raise my score to 7.

Authorsrebuttal2024-08-12

We thank the reviewer for their additional feedback. We will include in the final version of the paper a specific discussion of non-convex and discrete constraints (also addressing comments by reviewer `yA4S`), and a more detailed explanation and justification of the tuning of DC3. We will also include a figure comparing the evolution of the Lagrangian bound obtained by DLL and DC3 during training.

Reviewer vMXP2024-08-12

Response to rebuttal

Thank you to the authors for their detailed response to my questions and comments, as well as the general comments that they have provided to the overall feedback from all reviewers. I think that I agree with the authors about practical problems vs. edge cases, and it sounds like computational tools and findings in practice can achieve a high performance with this framework. I am happy to change my rating to a 7.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC