Winner Takes It All: Training Performant RL Populations for Combinatorial Optimization

Applying reinforcement learning (RL) to combinatorial optimization problems is attractive as it removes the need for expert knowledge or pre-solved instances. However, it is unrealistic to expect an agent to solve these (often NP-)hard problems in a single shot at inference due to their inherent complexity. Thus, leading approaches often implement additional search strategies, from stochastic sampling and beam search to explicit fine-tuning. In this paper, we argue for the benefits of learning a population of complementary policies, which can be simultaneously rolled out at inference. To this end, we introduce Poppy, a simple training procedure for populations. Instead of relying on a predefined or hand-crafted notion of diversity, Poppy induces an unsupervised specialization targeted solely at maximizing the performance of the population. We show that Poppy produces a set of complementary policies, and obtains state-of-the-art RL results on four popular NP-hard problems: traveling salesman, capacitated vehicle routing, 0-1 knapsack, and job-shop scheduling.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ybFK6/10 · confidence 4/52023-06-26

Summary

This paper presents a multi-decoder (population) neural network structure and training method to solve the combinatorial optimization problem. In particular, the paper presents a model update method in which multiple decoders can specialize in different types of problem instances. The method proposed in this paper shows promising results in TSP, CVRP, KP, and JSSP experiments.

Strengths

**S1.** The idea of agent population and the training method used for building it are novel. **S2.** While the idea is simple and clear, the accuracy improvement by the proposed method is remarkable. When performing inference in a fast and simple form(i.e. without iterative searching for inference), Puppy showed superior results compared to POMO 16 samples or POMO 16 ensemble. **S3.** The method proposed in the paper is applicable to various types of CO problems. **S4.** The idea of applying the multi-agent method of this paper to the existing POMO training method is excellent.

Weaknesses

There seems to be no major flaw in the methods presented in the paper. However, there are two opinions related to the TSP/CVRP experiment results. **W1.** Recent studies that have shown good results in TSP and CVRP experiments are omitted as baselines in Tables 1 and 2, for example [1, 2, 3]. Especially in the case of CVRP experiments, EAS[1], DPDP[2] and SGBS[3] have outperformed LKH3. **W2.** HGS [4,5] shows better performance than LKH3 as a meta-heuristic algorithm. I recommend that paper authors consider adding HGS as a meta-heuristic baseline in Table 2. **References** [1] Andre Hottung, et al. Efficient active search for combinatorial optimization problems. International Conference on Learning Representations, 2022. [2] Kool, Wouter, et al. Deep policy dynamic programming for vehicle routing problems. Integration of Constraint Programming, Artificial Intelligence, and Operations Research, 2022. [3] Jinho Choo, et al. Simulation-guided Beam Search for Neural Combinatorial Optimization. Advances in Neural Information Processing Systems 35, 2022. [4] Thibaut Vidal, et al. A hybrid genetic algorithm for multidepot and periodic vehicle routing problems. Operations Research, 2012. [5] Thibaut Vidal. Hybrid genetic search for the cvrp: Open-source implementation and swap* neighborhood. Computers & Operations Research, 2022.

Questions

**Q1.** In Algorithm 1, reward of the result of the second best agent $R(\tau_i^{**})$ was used as the baseline(169\~171), but Algorithm 2 seems to have used the POMO shared-baseline of the result of the best agent as the baseline(Algorithm 2, 7\~9). In other words, parameter update by applying the POMO training algorithm only for the best agent for each instance. Is it correct? Please provide a more detailed explanation of the parameter update in Algorithm 2. **Q2.** How to determine the appropriate number of populations? Are there criteria, considerations, etc. for the decision? **Q3.** Can you show the change in accuracy, training time, inference time, etc. according to the change in the number of population? **Q4.** Algorithm 1 Input - Missing 'H' in 'number of training steps H'.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer 51fw5/10 · confidence 4/52023-07-04

Summary

This paper proposes a construction method that learns a population of agents to improve the exploration of the solution space of combinatorial optimization problems. Experiments demonstrates that the proposed method improves the solving efficiency of four popular NP-hard problems.

Strengths

1. The paper is well-written and easy to follow. 2. Experiments demonstrate that the proposed method achieves state-of-the-art RL results on four popular combinatorial optimization problems.

Weaknesses

1. The motivation is unconvincing. The authors provide a motivating example of a naïve decision-making environment in Figure 1. However, it would be more convincing if the authors could provide an example of combinatorial optimization problems, as this paper aims to develop effective methods to solve combinatorial optimization problems. 2. The idea of learning a population of agents with diverse policies is incremental, as [1] has proposed a similar method. 3. The authors claim that their proposed method can produce a set of complementary policies. However, there is no proof for this claim. Thus, it would be more convincing if the authors could provide theoretical and/or empirical proof for this claim. 4. The experiments are insufficient. (1) It would be more convincing if the authors could evaluate the proposed method on combinatorial optimization problems with larger sizes, such as TSP with 1000 cities. (2) Some important baselines are missing. First, the authors may want to compare their method with MDAM [1] on the packing problems. Second, the idea of learning a population of agents is similar to learning an ensemble of agents. Thus, the authors may want to compare their method with baselines with the ensemble learning trick. (3) It would be better if the authors could provide a time analysis of training cost, as training a population of agents may take much longer time than that of baselines. [1] Xin, Liang, et al. "Multi-decoder attention model with embedding glimpse for solving vehicle routing problems." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 13. 2021.

Questions

Please refer to Weaknesses for my questions.

Rating

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

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer wqBo8/10 · confidence 4/52023-07-06

Summary

The paper proposes a new training procedure that allows to train a diverse set of policies for solving combinatorial optimization problems. Most existing approaches for these problems train a single policy/agent and aim to construct a solution in a single shot (or by sampling multiple solutions). In contrast, the authors propose to use a population of agents that is trained such that a diverse set of solutions can be obtained by solving a problem instance by each population member. To this end, the paper proposes a new training objective that aims to increase the maximum reward over all K population members. This is implemented by backpropagating only through the best performing agent for each instance during the training phase. Experimentally, the authors show that this approach is indeed able to learn a set/population of complementary policies. Furthermore, the results demonstrate that the performance of a population trained via the proposed approach is clearly superior to a population in which all agents are trained independently in standard-fashion.

Strengths

- While other papers have proposed to train a population of agents, the proposed training procedure (that always only trains the best agent of the population per instance) is novel. - The proposed methods show a significant performance improvement on the considered problems. For example, on the traveling salesperson problem the method finds solutions with a gap to optimality of 0.07% (in comparison to a gap of 0.27% of the state-of-the-art POMO method). To the best of my knowledge, the proposed method is currently the best neural combinatorial optimization method for quick solution generation (i.e., without a search component). - The method succeeds in training a population of complementary agents. The authors show experimentally that the average solution quality decreases during the training while the best performance over all agents increases. This means that the agents successfully specialize on specific strategies. - The authors evaluate their method on 4 different combinatorial optimization problems which demonstrates the wide applicability of the technique. On all problems, the method significantly improves the performance. - The considered problem of learning heuristics for combinatorial optimization problems has gotten a lot of attention in the literature and is a very promising research area. - The proposed method is simple and can be easily applied to other neural construction methods. Thus, it is likely that other researchers will use the proposed technique in their work. - Overall, the paper is well written and clearly organized.

Weaknesses

- Overall, the additional training of a population with the proposed method is quite resource intensive because only one of the K rollouts is used for backpropagation. In some settings, the additional resources needed for the population training phase might not be worth the obtained performance improvements.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Contribution

3 good

Limitations

- The paper does not discuss limitations of the proposed method.

Reviewer Ym2T5/10 · confidence 4/52023-07-26

Summary

The paper proposes that populations of agents can produce better results than using single agents. This leads to a new "Poppy" algorithm which performs policy gradient updates only on the agent which produced the highest reward. This is then applied to combinatorial optimization problems using attention-based architectures to demonstrate fast and competitive results.

Strengths

* The proposed "Poppy" algorithm appears to be a novel policy gradient variant. I do wonder however, if such an algorithm can be extended to general RL settings, rather than only combinatorial optimization. * Overall the paper is a clean read and presentation is strong. I did not have any issues with understanding. * Experimental results are reasonable. The method demonstrates good performance and low inference costs.

Weaknesses

The core issue of the proposed method, is that there is no explicit "diversity" objective being optimized, which makes this a theoretical weakness. However, the paper does defend against this through multiple explanations (e.g. agents will tend to specialize) and empirical evidence (e.g. best agent improves which population worsens).

Questions

* Could the proposed Poppy RL algorithm be more generally applicable to any MDP? If so, which problems (outside of combinatorial optimization) do you think can be most benefited? * Following up on my weakness section - In which types of MDPs/cases could "diversity" behavior not appear from Poppy training, since diversity is not an explicitly optimized objective?

Rating

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

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

Limitations section appears comprehensive. No issues.

Reviewer ybFK2023-08-14

I appreciate the authors addressing my concerns with detailed and appropriate responses, as well as conducting additional experiments. However, I will keep my rating for this paper unchanged. I have no further questions.

Area Chair YoAd2023-08-16

AC note: Please engage with authors

Hi Reviewer 51fw, Please engage with the authors. They have put in a significant amount of effort to respond to the concerns, and improve their work. I'd encourage you to respond asap and give the authors an opportunity to continue improving their work!

Reviewer 51fw2023-08-17

Thanks for the kind reminder

Dear AC Thank you for the kind reminder. I have read the authors' rebuttal and the other reviews. I am engaging with authors now.

Authorsrebuttal2023-08-17

We thank the reviewer for the continued discussion. We regret that concerns still remain, however we remain confident that these points can be satisfactorily addressed and do not represent significant issues with our work. Whilst we hoped our original responses would clarify these points, we are more than happy to address them further. However, **we would kindly ask that the Reviewer clarify specifically why they do not feel our responses were sufficient** such that we can continue to address remaining questions. ### Concern 1 We are uncertain about the Reviewer's specific concerns. Is the question centered on the rationale for having a diverse policy set rather than a single policy when addressing NP-hard CO problems? Or does the Reviewer find the toy example in Figure 1 overly simplistic to illustrate this point? If it is the latter, we are open to substituting the maze with a more intricate routing problem. The overarching argument remains valid, should the Reviewer prefer this change. However, we wish to highlight that Poppy's robust empirical performance is our principal contribution. Our goal with Figure 1 and the related discussion was to elucidate the intuition behind proposing the Poppy objective in a simplified setting. Even if there are suggestions for alternative motivating examples or potential for deeper theoretical insights, we maintain that Poppy's demonstrated effectiveness justifies our contribution. ### Concern 3 We are confident that our results demonstrate that Poppy produces a set of complementary policies. We'll outline the evidence, and if the reviewer remains unconvinced, we'd value specific feedback. Should our manuscript lack clarity on this topic, we're open to making necessary revisions. By "a complementary set of policies", we mean: 1. Individual policies in the population are most performant on different sub-distributions of problem instances. 2. As a collective, the policies encompass the entire problem instance distribution, ensuring that the performance of the best policy for any sampled problem instance remains strong. **Evidence for (1)** Removing any agent from the population reduces the overall population-level performance, as seen in Fig 4's bottom-right (Appendix D.1). This signifies that every agent is optimal for a specific sub-distribution of problems. It's important to note that these contributions are computed with respect to greedy rollouts of each policy, and so cannot be attributed to stochasticity in the agent performance. Fig. 3 illustrates the effects of specialization. As the population size grows, the expected performance of a randomly selected agent on a random problem gets worse. However, the performance of the best-suited agent for a random problem is improved. This highlights that when each agent has a narrower target sub-distribution, it learns an even more specialized policy, which is even better (worse) for problem instances in (out) of the target sub-distribution. This is in contrast with a simple policy ensemble, for which the average agent performance would remain the same regardless of the population size. **Evidence for (2)** In a head-to-head comparison of Poppy's collective performance against POMO – either sampling a single POMO agent N times, or training an ensemble of N POMO agents – Poppy consistently surpasses both benchmarks at the same cost (Tables 1, 2 and 3). This demonstrates that the set of policies learned by Poppy, as a collective, excels across all problem instances. ### Concern 4 **4.1**: As described in our original response, we agree that scaling to extremely large CO problems is an important challenge but note that it is a distinct research topic. Baseline methods to which Poppy is compared (POMO, EAS, MDAM, LIH, NeuRewriter, NLNS, 2-Opt-DL, CVAE-Opt, DPDP) tackle the same order of problem sizes as our work. This ensures fair comparison and our leading performance in these problem settings (see Tables 1, 2 and 3) is not questioned by any reviewers. If the Reviewer’s belief is that scaling up to TSP1000 is mandatory for all research in this domain, we would respectfully disagree and note this is a standard not applied to prior works. **4.2.1**: The requested MDAM baseline for KP and JSSP is shown to be significantly outperformed by other RL methods on TSP and CVRP (Tables 1 and 2). Given the absence of published results for MDAM on these problems and the impracticality of re-implementing all existing neural CO methods, we prioritized the strongest baselines: POMO+sampling and POMO+ensembling. If the Reviewer does not agree this is reasonable we would welcome specific inquiries. **4.2.2**: The requested baseline was included in our original submission, so we believe this issue should be considered resolved.

Area Chair YoAd2023-08-16

AC Note: Please engage with the authors

Hi Reviewer Ym2T, Please engage with the authors. They have put in a significant amount of effort to respond to the concerns, and improve their work. I'd encourage you to respond asap and give the authors an opportunity to continue improving their work!

Reviewer Ym2T2023-08-16

Keeping current score of acceptance

I thank the authors for their time writing their response. I do think the paper has adequately defended its position on the diversity issue, and has achieved good results. The proposed method indeed is interesting, and it appears to be a more general technique than just for combinatorial optimization (although its generality is not explored in this paper). At the moment, the paper focuses mostly on combinatorial optimization cases however, and so the impact of the paper may be limited. Thus I am fine with lukewarm acceptance at the moment.

Authorsrebuttal2023-08-17

We appreciate the Reviewer's feedback and are thankful our work's merit isn't in doubt. The Reviewer comments on a possible narrow appeal due to our focus on CO problems. We respectfully differ in this view. Combinatorial Optimization is a significant and long-standing research area, with related papers frequently presented at top venues like NeurIPS. For reference, we've listed several recent NeurIPS publications on neural CO at the end of our response. To further address the Reviewer’s concern, we're open to expanding our introduction and related work sections to underscore real-world applications of CO problems, such as routing, scheduling, and chip placement. However, we fully appreciate that gauging the impact of specific topics can be subjective and simply wished to take this chance to share our perspective. We're grateful for the reviewer's constructive feedback and their positive evaluation of our research. **NeurIPS 2022** * Malherbe et al. “Optimistic Tree Searches for Combinatorial Black-Box Optimization”. * Choo et al. “Simulation-guided Beam Search for Neural Combinatorial Optimization”. * Kim et al. “Sym-NCO: Leveraging Symmetricity for Neural Combinatorial Optimization”. * Wang et al. “Unsupervised Learning for Combinatorial Optimization with Principled Objective Relaxation”. **NeurIPS 2021** * Kim et al. “Learning Collaborative Policies to Solve NP-hard Routing Problems”. * Ma et al. “Learning to Iteratively Solve Routing Problems with Dual-Aspect Collaborative Transformer”. * Wang et al. “A Bi-Level Framework for Learning to Solve Combinatorial Optimization on Graphs”. * Kwon et al. “Matrix encoding networks for neural combinatorial optimization”.

Reviewer 51fw2023-08-17

Thanks for the authors’ rebuttal

Thanks for the authors' response. However, my major concerns 1, 3, and 4 have not been properly addressed. Thus, I would lean toward rejection given the current status of my communication with the authors. 1. (Concern 1 in weaknesses) The authors do not provide a convincing example for the motivation. 2. (Concern 3 in weaknesses) It is unclear to me why Fig 3 demonstrates that Poppy can produce a set of complementary policies. 2. (Concern 4 in weaknesses) My concerns 4.1 and 4.2 have not been properly addressed.

Reviewer wqBo2023-08-21

Thank you for your response. I keep my very high rating and believe that the paper should be accepted.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC