acceptodds
Under review as a conference paper at ICLR 2027

Actor Score Composition and Refinement for Multi-Objective Flexible Job Shop Scheduling

Abstract

Multi-objective flexible job shop scheduling has attracted growing research interest due to the need to balance competing objectives in production. Current neural approaches use preferences to condition model features, parameters, or computation paths, enabling a single model to construct schedules across different trade-offs. However, generating high-quality schedule sets is challenging because a single model needs not only to learn assignment and sequencing rules required across preferences but also to continually adapt its decisions to changing machine workloads and remaining operations. To address this challenge, we propose Actor Composition with Preference Correction (AC-PC), which constructs preference-conditioned scheduling policies by composing and refining independent actor scores. Specifically, a shared conditional encoder represents the current scheduling state at fixed preference anchors, and independent actors produce action scores that are combined using weights determined by the requested preference. Moreover, preference correction uses the requested preference and mixed actor features to apply an action-specific adjustment to the composed scores. To evaluate AC-PC, experiments are conducted on same-size synthetic instances, larger unseen instances, and public benchmarks. The results show that AC-PC outperforms baselines in most cases.

Primary area:
optimization
Keywords:
Multi-objective Combinatorial Optimization, Flexible job shop scheduling, Deep reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_feAiOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies multi-objective flexible job shop scheduling, aiming to generate schedules that reflect different objective preferences using a single neural policy. Building on the DCAN framework, the authors propose AC-PC, which uses a shared conditional encoder to encode the scheduling state at fixed preference anchors. Independent actors score candidate operation–machine pairs, and their scores are combined according to the requested preference. A preference correction network uses the mixed hidden features and the requested preference to refine candidate action scores. The entire policy is trained jointly using PPO.

The contributions include preference-conditioned composition of independent action scores, action-level preference correction, and systematic experimental validation of the design. The experiments cover different numbers of objectives, instance sizes, public benchmarks, and other scheduling problems. They demonstrate hypervolume improvements over DCAN in most tested settings and examine the roles of individual components through ablations, actor-parameter-matched comparisons, and PC-on/off analyses.

Critical strengths
  1. The paper proposes a clear and practical preference-conditioned design at the action-scoring level. It combines independent actor scores at fixed preference anchors so that different scoring branches jointly contribute to objective trade-offs, then uses the current state and the requested preference to correct the score of each candidate action. The design clearly separates score composition and score refinement, with both components trained jointly through a single policy, providing a well-structured approach to multi-objective scheduling that is straightforward to implement.

  2. The experimental evaluation is generally comprehensive, covering two-, three-, and four-objective settings, different instance sizes, public benchmarks, and other scheduling problems such as FFSP and JSSP. In the same-size two- and four-objective experiments, AC-PC outperforms DCAN in all 16 two-objective comparisons and all 8 four-objective comparisons, with relative HV improvements of 10.26%–28.01% on the four-objective tasks. Transfer experiments on larger instances and public benchmarks further demonstrate the method's generalization ability, while component ablations and actor-parameter-matched comparisons help explain the sources of performance improvements. The authors also report cases where the method does not outperform the baselines, making the scope of the experimental conclusions reasonably clear. Overall, these results provide reasonably convincing support for the method's effectiveness and practical value.

Critical weaknesses

The submitted code omits parts of the multi-actor implementation, affecting the reproducibility of the experiments and making it more difficult to understand the full model implementation.

Questions for authors
  1. In the scheduling illustration in Figure 1, operation O22O_{22} starts before its predecessor O21O_{21} has finished. Does this violate the job precedence constraints? Could the authors check and clarify this?

  2. Could the authors provide a more detailed implementation, particularly for the multi-actor component, along with the corresponding configurations and instructions for running the code, to help readers understand the model implementation and reproduce the experimental results?

3 comments
AuthorsOpenReview ↗

Response on Figure 1 and the complete multi-actor implementation

Thank you for your positive assessment of our work and your helpful comments. References below refer to the revised manuscript.

Q1. Precedence in Figure 1 (Section 4). We have corrected the timing bars in Figure 1 so that O₂₂ starts at or after O₂₁ finishes. The scheduling environment enforces Sᵢⱼ=max(Cᵢ,ⱼ₋₁,Aₘ), using the predecessor's completion time and the machine's availability to determine each operation's start time.

W1/Q2. Implementation details. As stated in the Reproducibility statement, we will open-source the complete implementation immediately upon acceptance, including the multi-actor model, training/evaluation code, configurations, dependencies, and running instructions.

Section 4 describes the model: the shared encoder processes the scheduling state at each fixed preference anchor. Each actor has separate parameters and outputs action scores and hidden features. Preference-dependent weights combine these outputs; PC uses the mixed features and requested preference to refine each candidate's score before feasible-action masking and softmax. The encoder, actors and PC are trained jointly through the composed policy using PPO.

I thank the authors for their rebuttal, which satisfactorily addresses the concerns raised in my review. I therefore remain my score.

At the same time, I recognize that my review may not fully capture the concerns highlighted by other reviewers. I would welcome their perspectives on whether the rebuttal adequately resolve these issues, and I remain open to revisiting my recommendation in light of the discussion.

AuthorsOpenReview ↗

Thank you for your follow-up

Dear Reviewer feAi,

Thank you for reviewing our response and for your support of our work. We are glad that our explanations address your concerns, and we appreciate your helpful comments and time throughout the review process.

Best regards, The Authors

Reviewer_8h3BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Actor Composition with Preference Correction (AC-PC), a deep reinforcement learning method for multi-objective flexible job shop scheduling. A shared encoder represents scheduling states at fixed preference anchors, and independently parameterized actors provide separate action-scoring mappings whose outputs are combined according to the requested preference. A preference correction module further adjusts these scores using the preference and mixed actor features. Experiments on synthetic instances, larger unseen instances, and public benchmarks show that the proposed method outperforms baseline methods in most evaluated settings.

Critical strengths
  1. The motivation is clear, and the architecture closely follows the stated challenges. Independent actors provide separate scoring mappings, while PC enables action-specific adjustments conditioned on the preference and scheduling state. This makes the intended roles of the components easy to understand.

  2. The evaluation spans multiple objective combinations, instance sizes, and public benchmarks, including transfer without retraining. Improvements over DCAN in most settings provide evidence of practical value beyond the training instances.

Critical weaknesses
  1. The anchor and Bernstein-basis choices appear somewhat heuristic, with limited justification for using a quadratic basis for two objectives but a linear one for three, and no sensitivity analysis. Table 4 suggests that most gains come from independent actors, while PC provides marginal or inconsistent improvements and even underperforms AC on the 20×1020\times10 three-objective task. The lack of a separate AC-only ablation for four objectives, where the largest gains occur, further weakens the evidence for PC's effectiveness.

  2. The reported runtimes use different hardware platforms, as acknowledged in Appendix A, and therefore do not provide a controlled efficiency comparison. AC-PC also requires multiple anchor-conditioned encoder and actor evaluations per decision. Although preference counts and sampling budgets are matched, the lack of comparisons on the same hardware under comparable computational budgets leaves its solution-quality advantage at comparable computational cost unclear.

Questions for authors
  1. What motivates the choice of three anchors and the quadratic Bernstein basis for two objectives versus the linear basis for three, beyond computational considerations? Would changing the number or placement of anchors, or the basis functions, substantially affect performance?

  2. Given that AC already captures state and preference information, could PC's additional correction be redundant? How can the authors demonstrate that the improvements, particularly for four objectives, are not primarily attributable to AC alone?

  3. Could the authors evaluate DCAN and AC-PC on the same hardware under consistent timing protocols, reporting training and inference costs separately? Does AC-PC retain its performance advantage under comparable computational budgets?

2 comments
AuthorsOpenReview ↗

Response on anchor design, four-objective ablations, and computational cost

Thank you for the focused questions. References below refer to the revised manuscript.

W1/Q1. Anchors and basis. Section 4.2.1 defines endpoint and midpoint anchors for pure and balanced two-objective preferences, with quadratic Bernstein weights. For three objectives, simplex vertices represent the pure objectives, and barycentric weights interpolate their scores directly. Both bases are smooth, nonnegative and sum to one. Table 19 in Appendix N compares two-objective variants trained from scratch with PC for 1500 updates (seed 300); tests use 100 SD1 20×10 instances, 101 preferences and 10 samples/preference. Beta τ=4 keeps the three anchors and normalizes (w₁⁴,6w₁²w₂²,w₂⁴).

Actors / basisGreedy HVSampling HV
2 / linear0.81310.8232
Original 3 / quadratic0.81380.8251
5 / quartic0.81350.8245
3 / Beta τ=40.81300.8246

The three-actor quadratic default has the highest HV in both modes among the actor-count and basis variants above; five actors add parameters without improving HV. Beta demonstrates that the weighting family can be replaced within the same framework.

To isolate anchor location, we retrain the three-actor model with only the middle encoder anchor changed from (0.5,0.5) to (0.25,0.75), retaining endpoint anchors, quadratic Bernstein weights, PC and no actor hidden-layer residual. Training uses seed 300 and 1500 updates, with checkpoint selection by validation HV; testing follows the same protocol above:

Middle anchorGreedy HVSampling HV
(0.5,0.5), original0.81380.8251
(0.25,0.75), retrained0.81200.8216

The default midpoint achieves higher HV in both greedy and sampling modes than the shifted anchor, supporting our default anchor choice.

W1/Q2. PC's role. Actor composition (AC) already uses state and preference information, but combines branch scores linearly with preference-only coefficients. Section 4.3 describes how PC models nonlinear interactions in the mixed features and preference to refine each candidate's composed score. Table 7 in Section 5.7 evaluates four objectives on 100 SD1 20×10 instances, with 120 preferences and 10 samples/preference:

MethodGreedy HVSampling HV
AC, trained without PC0.61180.6244
Original AC-PC, re-evaluated0.61260.6282
Same AC-PC checkpoint, PC off at test0.59890.5762

Adding PC improves HV over AC trained without PC in both modes. The larger drop from removing PC at test time shows the trained policy's reliance on the correction.

W2/Q3. Quality and cost. Appendix P, Table 22 reports scheduling-loop times on one RTX 5090 for makespan–costs (two objectives) and four objectives:

Objectives / methodGreedy s/instanceSampling s/instance
Two / DCAN2.950027.9600
Two / AC-PC3.490034.3700
Four / DCAN4.689348.1269
Four / AC, trained without PC5.791059.1955
Four / AC-PC5.801058.8081

Matched two-objective training including validation takes 3.9300 h for AC-PC versus 2.9800 h for DCAN. Appendix P gives the full protocols.

Table 8 in Section 5.7 compares HV under equal per-instance scheduling-loop time budgets:

Budget (s)Two: DCANTwo: AC-PCFour: DCANFour: ACFour: AC-PC
50.80700.81030.55030.58630.5930
100.81010.81390.55230.60820.6090
200.81520.81850.55790.61370.6153
400.81740.82190.56120.61830.6202

AC-PC achieves higher HV than DCAN at every tested budget for both objective settings, and exceeds AC trained without PC at all four-objective budgets. Timing includes policy evaluation, action selection and environment updates; only completed blocks of 16 schedules count, with the same preference order and one greedy sweep before sampling.

PC adds little inference cost relative to AC: four-objective sampling takes 58.8081 versus 59.1955 s/instance, while HV increases from 0.6244 to 0.6282.

Thank you to the authors for the detailed rebuttal and careful revisions. The additional experiments and explanations have largely addressed my previous concerns.

My only remaining reservation concerns the relatively modest performance gains provided by the PC module. In contrast, I find the idea of independent actors and score composition interesting, with clearer performance improvements. Moreover, the additional parameter-matched comparisons and experiments with separately trained DCAN policies, provided in response to another reviewer's questions, suggest that the benefits of AC cannot be explained simply by increasing the parameter count.

Overall, I consider this paper to be borderline between scores of 2 and 3, and I am open to either an acceptance or rejection decision. If the other reviewers also find that their main concerns have been adequately addressed, I would be willing to reconsider my rating.

Additionally, I would appreciate it if the authors could highlight the revised and newly added content in the manuscript using different colors or other clear markings. This would make it easier for reviewers to identify the specific changes.

Reviewer_6vtwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Actor Composition with Preference Correction, or AC-PC, for preference-conditioned multi-objective flexible job shop scheduling. Its first contribution is a shared conditional encoder that represents the partial schedule at fixed preference anchors, plus independent actors that learn separate scoring rules for candidate operation and machine pairs. Composition weights depend only on the requested preference, so one policy can score actions across preferences without retraining. Its second contribution is a gated preference correction that refines the composed scores with an action-specific adjustment, and it vanishes at preference endpoints or simplex vertices so the base scores are preserved there. All branches train jointly through one composed PPO policy with feasible-action masking. The third contribution is an evaluation across training sizes, larger unseen instances, public benchmarks, four objectives, transfer without retraining, ablations, and a parameter-matched comparison. Results show AC-PC exceeds DCAN in most tested settings, and the benefits vary by task and decoding mode. Limitations are that weighted-sum training cannot cover unsupported Pareto regions, and each decision needs multiple anchor-conditioned encoder evaluations.

Critical strengths
  1. The multi-objective FJSP is a more realistic scheduling problems in the manufacturing scheduling domains. Most of the papers are targeting the single objective FJSP.
  2. The targeted objectives, such as Makespan、Tardiness、Flowtime、and Costs are important metrics for measuring the efficiency of manufacturing systems.
  3. It is smart to saperate state-action learning and preference learning. I like this idea.
  4. The paper is well-written and easy to follow.
Critical weaknesses
  1. The paper lacks Inverted Generational Distance(IGD)as primary metrics for multi-objective optimization
  2. This paper is built on top of DCAN, which harms its novelty to some extent. The key difference is that the actors have saperate parameters to learn, and the preference is treated as weight to adjust the sampling probs (Bernstein-basis score/feature composition).
  3. From results in Table4, the preference correction does not contribute that much performance boost to the whole framework. The main gain is from saperate parameters.
Questions for authors
  1. I don't understand the second challenge in the Introduction. Can you elaborate a bit more?
  2. Why a shared encoder can learn an anchor-conditioned representations for difference objectives?
  3. If the saperate parameters are the key attribute to the performance gain against DCAN, what if we give DCAN saperate policies and how its performance will be?
1 comment
AuthorsOpenReview ↗

Response on conditional representations, PC, and independent policies

Thank you for your questions. References below refer to the revised manuscript.

W1. Inverted generational distance (IGD). Appendix O, Table 21 reports the following results on 100 SD1 20×10 instances. For each objective setting, methods share an empirical nondominated reference front pooled across methods and both decoding modes, with per-instance objective-wise min/max normalization. AC denotes actor composition trained without PC; AC-probability uses probability mixing throughout training.

Objectives / methodGreedy IGDSampling IGD
Two / DCAN, newly trained0.04350.0173
Two / three regional DCANs0.14040.0789
Two / original AC-PC0.04110.0159
Three / DCAN0.07450.0311
Three / AC-PC0.07140.0274
Three / AC-probability0.20950.1648
Four / DCAN, newly trained0.14150.0820
Four / AC, trained without PC0.10600.0565
Four / original AC-PC0.09860.0409
Four / same AC-PC checkpoint, PC off at test0.13020.1291

AC-PC has lower IGD than DCAN in both modes for two, three and four objectives. For four objectives, it also improves over AC and the same AC-PC checkpoint with PC disabled at test time. Section 5.7 summarizes these results.

W2. Contribution relative to DCAN. Building on DCAN's conditional encoder and PPO, AC-PC separates anchor-specific scoring from preference-dependent composition and candidate-wise refinement. Separately parameterized actors learn scoring mappings at fixed anchors; their scores are composed for the requested preference and refined by PC. These components train jointly through the resulting policy, linking specialized scoring with preference-conditioned decisions.

Table 5 compares three independent actors (20,547 actor parameters) with a widened shared actor (20,709) on the 20×10 tasks, both without PC and using the same anchors and training/evaluation protocol. Independent actors improve both decoding modes on both two-objective tasks. Three-objective greedy/sampling differences are +0.0007/−0.0033.

The four-objective comparison in Section 5.7, Table 7 matches total parameters exactly: both variants have 453,069 (446,220 trainable), retain the same four anchor-conditioned encoder passes, omit PC, and train for 1500 updates with seed 300 and the same PPO settings. Independent actors improve greedy/sampling HV by 0.0150/0.0093. The earlier two-anchor ablation in Appendix N also favors independent actors in all 16 task/size/decoding comparisons. At 20×10, sampling HV rises from 0.8202 to 0.8262 for makespan–costs and from 0.8309 to 0.8397 for tardiness–costs.

W3. PC's increment. PC refines each candidate's composed score through nonlinear interactions in the mixed actor features. With actor architecture and decoding budget fixed, adding PC improves sampling HV in seven of eight two-objective and three of four three-objective comparisons (Table 4).

Q1. Second challenge. Section 1 identifies the need to adjust the composed scores differently for different operation–machine assignments. AC's branch scores already depend on the scheduling state, but all candidates use the same preference-dependent mixing coefficients. Section 4.3 describes how PC uses each candidate's mixed features and the requested preference to compute a nonlinear adjustment to its composed score.

Q2. Shared encoder (Section 4.1). Eθ(s,vₖ) shares parameters θ but receives different anchor inputs vₖ. Learned operation/machine preference projections inject these anchors into the encoder, so the same scheduling state can produce different representations at different anchors. PPO gradients through the corresponding actors jointly train these conditional representations.

Q3. Separate DCAN policies (Appendix N). For makespan–costs on 100 SD1 20×10 instances, with 101 test preferences and 10 samples/preference, we compare DCAN and AC-PC (Table 1) with three independently trained regional DCAN policies:

MethodGreedy HVSampling HV
DCAN0.80830.8200
Three independent regional DCANs0.80790.8214
AC-PC0.81380.8251

Each regional policy has its own encoder, actor and critic, trained for 1500 updates (seed 300) with uniform makespan weights in [0,.25], [.25,.75] or [.75,1]. Each test preference is routed to its corresponding policy. The regional policies slightly improve sampling HV over DCAN but give no greedy gain; AC-PC remains higher in both modes.

Reviewer_zJhtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper solves Multi-Objective Flexible Job Shop Scheduling (MOFJSP) by extending the actor architecture of DCAN [1]. The proposed AC-PC combines the action logits produced by multiple independent actor heads through preference-dependent weights and further adjusts them using a Preference Correction MLP. However, the ablation results suggest that Preference Correction provides only limited and inconsistent benefits.

The motivation is not well established. The first challenge starts with O–M pair selection, which is a fundamental decision in FJSP, and suddenly turns to different preferences. The paper does not clearly explain why this challenge specifically arises from different preferences. Moreover, the proposed architecture does not appear to introduce a new mechanism for addressing the second challenge of continual adaptation to changing scheduling states.

Therefore, I would like to rate this paper around 1~1.5.

Critical strengths

The appendix is comprehensive and provides extensive experimental results and implementation details.

Critical weaknesses
  1. The proposed method is mainly an extension of DCAN's actor architecture, introducing independent actor heads and a Preference Correction MLP. The technical modifications are relatively simple, and the ablation results show that Preference Correction provides negligible improvements in several cases and even degrades performance in others.

  2. The presentation lacks a coherent motivation for the proposed architecture. In particular, the two challenges identified in the Introduction are not clearly justified, and their connections to the proposed technical components remain unclear.

Questions for authors
  1. Why are exactly three actors used? Could alternative weighting schemes, such as a Beta-distribution-based scheme, be considered for controlling the contributions of the three actors?

  2. Could the authors evaluate the three actors separately on different objectives and summarize their individual objective values in a table?

  3. How do the three actors evolve during joint training? Please provide the training curves of their action entropy across PPO updates to investigate how actor diversity changes throughout training.

  4. Why is Logit Aggregation adopted? Could the authors include an ablation comparing Logit Composition with Probability-level Mixture?

  5. The three actors are jointly trained rather than independently pretrained as objective-specific expert policies. Could the authors compare against expert-based approaches? For example, PA-MoE [2] has already explored preference-specific experts and preference-aware gating for multi-objective combinatorial optimization.

  6. Makespan and Flowtime are not necessarily conflicting objectives, which raises concerns about the motivation for the three- and four-objective formulations. Moreover, at 20×10, AC-PC performs worse than DCAN in the three-objective sampling experiment but achieves substantial improvements in the four-objective setting. Could the authors explain this inconsistency?

  7. The baseline comparisons are insufficient. Please include at least two relevant methods from [2]–[5] for a more comprehensive evaluation.

  8. (Minor) The generalization experiments on FFSP and JSSP in Appendices K and L are not adequately discussed in the main text. Please briefly summarize these results.


[1] DCAN: Neural Multi-Objective Combinatorial Optimization for Flexible Job Shop Scheduling Problems. ICLR 2026.

[2] PA-MoE: Preference-Aware Mixture-of-Experts for Multi-Objective Combinatorial Optimization. OpenReview 2026.

[3] Neat-Weights: Rethinking Neural Multi-Objective Combinatorial Optimization via Neat Weight Embedding. ICLR 2025.

[4] POCCO: Preference-Driven Multi-Objective Combinatorial Optimization with Conditional Computation. NeurIPS 2025.

[5] WeCon: An Efficient Weight-Conditioned Neural Solver for Multi-Objective Combinatorial Optimization Problems. arXiv 2026.

2 comments
AuthorsOpenReview ↗

Response on actor specialization, composition, and correction

Thank you for the questions. References below refer to the revised manuscript.

W1. Contribution. Building on DCAN's encoder and PPO, AC-PC jointly learns anchor-specific scorers, composes their scores, and refines them with PC. Without PC, independent actors improve both decoding modes on both two-objective tasks with slightly fewer actor parameters (Table 5), and on four objectives at equal total parameter count (Table 7). With actor architecture and decoding budget fixed, PC improves sampling HV in seven of eight two-objective and three of four three-objective comparisons (Table 4).

W2. Architectural rationale. Figure 1 outlines the architecture detailed in Sections 4.1–4.5. Different preferences can favor different assignments. A shared anchor-conditioned encoder supplies features to independent actors, which learn separate scoring mappings. Preference weights combine their scores and hidden features for the requested trade-off. These weights are shared across candidates; PC uses each candidate's mixed features and preference to learn a nonlinear correction to its composed score. Masking and softmax yield one policy, through which PPO jointly trains the encoder, actors and PC.

Q1. Anchors and weights. In Section 4.2.1, two-objective pure/balanced anchors use quadratic Bernstein weights; three-objective vertex anchors use barycentric weights. Both bases are smooth, nonnegative and sum to one. With PC, the three-actor quadratic default has the highest greedy/sampling HV among tested count/basis variants (Appendix N, Table 19). Five actors do not improve HV; Beta weights provide an alternative basis.

Q2. Individual actors (Appendix M, Table 18). Evaluating branches directly without composition or PC shows that the two-objective midpoint trades makespan against cost, and each three-objective branch is best on its anchor objective.

Q3. Training dynamics. Normalized branch entropies change from ≈(1,1,1) to (0.4841,0.6814,0.4117); pairwise Jensen–Shannon (JS) divergence grows from below 4.5000×10⁻⁶ to 0.3918–0.4217; final action disagreement is 82.5000%–90.0000%. A new three-objective run provides 39 checkpoints (seed 300, 1500 updates), evaluated on 40 fixed states from 10 validation instances. These trends show distinct policies emerging during training. Appendix M reports JS divergence and the entropy and action-disagreement curves (Figure 3). Section 5.7 summarizes these findings.

Q4. Logits versus probabilities. Without PC, logit composition gives a weighted geometric combination of branch probabilities, favoring agreement; arithmetic mixing can favor actions supported mainly by one branch. Training with probability mixing gives two-/three-objective sampling HV 0.8055/0.5883, below actor composition (AC) with logits at 0.8248/0.6315 (Table 4). Appendix N gives the equations and evaluation protocol.

Q5. Joint training and independent experts (Appendix N, Table 20). Without PC, three DCAN experts use 1,272,843 total parameters versus jointly trained AC's 444,828 (2.86×), changing greedy/sampling HV by −0.0005/+0.0018. Each expert has its own encoder, actor and critic and 1500 training updates (4500 total); AC shares one encoder and trains jointly for 1500 updates. Composition beats nearest-anchor routing with the same experts in both modes.

Q6. Three-/four-objective performance (Appendix N). AC-PC exceeds DCAN in seven of eight three-objective HV comparisons and all eight four-objective comparisons (four sizes, greedy/sampling; Tables 2/3).

Our interpretation is that the larger four-objective gains arise from the additional conflict introduced by earliness. Makespan and flowtime favor earlier overall completion and shorter job durations. For a job completed before its due date, earliness instead decreases when it finishes later. AC-PC adds a dedicated fourth actor for this competing priority and composes its scores according to preference. Separating these opposing priorities into distinct scoring mappings may be more beneficial than representing them in DCAN's single preference-conditioned scorer.

Q7. Neural multi-objective optimization baselines. We adapt WE-CA and POCCO-W to flexible job shop scheduling. Greedy/sampling HV is 0.7707/0.7780 for WE-CA, 0.7170/0.7237 for POCCO-W, and 0.8138/0.8251 for AC-PC. Both use 30,000 training trajectories and validation-HV selection; WE-CA uses PPO, POCCO-W retains its pairwise preference loss. See Appendix N.

Q8. Other scheduling problems. We train separate models for the flexible flow shop problem (FFSP) and job shop scheduling problem (JSSP). AC-PC improves all eight JSSP comparisons over DCAN (0.1100%–2.3700% relative HV). FFSP is close at 15 jobs/five stages and improves both modes at 20 jobs/four stages. See Section 5.7 and Appendices K/L (Tables 16/17).

Thank you to the authors for the clarification. My main concerns remain the limited methodological novelty and poor presentation. The proposed method is essentially an extension of the actor architecture in DCAN [1], and one of the introduced components makes only a limited contribution. Moreover, the paper suffers from poor logical organization and unclear presentation. Therefore, I maintain my original rating.


[1] DCAN: Neural Multi-Objective Combinatorial Optimization for Flexible Job Shop Scheduling Problems. ICLR 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response on component contributions, equal-budget performance, and stability

Thank you for highlighting component attribution, computation, and stability. References below refer to the revised manuscript.

W1/Q1. Four-objective contributions. Table 7 in Section 5.7 compares actor composition (AC) trained without PC, the original AC-PC, and the same AC-PC checkpoint with PC disabled at test time. Tests use 100 SD1 20×10 instances, 120 preferences and 10 samples/preference:

MethodGreedy HVSampling HV
AC, trained without PC0.61180.6244
Original AC-PC, re-evaluated0.61260.6282
Same AC-PC checkpoint, PC off at test0.59890.5762

Adding PC improves HV over AC trained without PC in both modes. The larger drop from disabling PC at test time shows the trained policy's reliance on the correction.

To isolate actor parameterization, we compare four separate actors with one wider shared actor. Both use the same four anchor-conditioned encoder passes, disable PC, and have exactly 453,069 total parameters (446,220 trainable). Both are trained for 1500 updates with seed 300 and the same PPO settings, and tested under the protocol above:

Actor parameterization (no PC)Greedy HVSampling HV
Shared actor, capacity matched0.59690.6152
Four separate actors0.61180.6244

At matched parameter count, separate actors improve greedy/sampling HV by 0.0150/0.0093.

W2/Q2. Runtime. Appendix P, Table 22 reports the following four-objective scheduling-loop times on one RTX 5090:

MethodGreedy s/instanceSampling s/instance
DCAN4.689348.1269
AC, trained without PC5.791059.1955
AC-PC5.801058.8081

PC adds little inference cost relative to AC: sampling takes 58.8081 versus 59.1955 s/instance, while HV increases from 0.6244 to 0.6282. For makespan–costs, AC-PC/DCAN sampling takes 34.3700/27.9600 s/instance; matched training including validation takes 3.9300/2.9800 h. Appendix P gives the full protocols.

Table 8 in Section 5.7 compares four-objective HV under equal per-instance scheduling-loop time budgets:

Budget (s)DCAN HVAC HVAC-PC HV
50.55030.58630.5930
100.55230.60820.6090
200.55790.61370.6153
400.56120.61830.6202

AC-PC exceeds both DCAN and AC at all four budgets. Table 8 also reports higher two-objective HV than DCAN at every tested budget. Timing includes policy evaluation, action selection and environment updates; only completed blocks of 16 schedules count, with the same preference order and one greedy sweep before sampling.

Q2. Training stability. We evaluate makespan–costs, tardiness–costs and four objectives with training seeds 300, 301 and 302. Each task uses the same 100 SD1 20×10 test instances across seeds, with 101/120 preferences for two/four objectives and 10 samples/preference. For four objectives, AC-PC seed 300 reuses the original checkpoint, re-evaluated under the same test protocol. We report mean HV ± sample standard deviation across seeds (denominator n−1), after averaging over instances for each seed:

TaskMethodGreedy HVSampling HV
Makespan–costsDCAN0.8019 ± 0.01230.8189 ± 0.0034
Makespan–costsAC-PC0.8144 ± 0.00150.8225 ± 0.0072
Tardiness–costsDCAN0.8122 ± 0.00770.8286 ± 0.0101
Tardiness–costsAC-PC0.8264 ± 0.00320.8339 ± 0.0081
Four objectivesDCAN0.5532 ± 0.00400.5686 ± 0.0067
Four objectivesAC-PC0.6114 ± 0.00520.6239 ± 0.0060

AC-PC achieves higher mean HV than DCAN in both decoding modes on all three tasks. For four objectives, it also achieves higher HV in both modes for all three paired seeds.

W3. Standalone implementation (Reproducibility statement). We will open-source the complete implementation immediately upon acceptance, including the multi-actor model, training/evaluation code, configurations, dependencies, and running instructions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.