acceptodds
Under review as a conference paper at ICLR 2027

PDScheduler: Persistent Dynamic Policy Decomposition for Flexible Job-Shop Scheduling

Abstract

Despite rapid progress in neural methods for the Flexible Job Shop Scheduling Problem (FJSP), existing paradigms face two key bottlenecks as instances grow: two stage approaches commit to machine assignments prematurely, rendering the downstream sequencing search prohibitively expensive, while autoregressive methods face decision horizons that grow with problem scale, making long horizon reasoning difficult and inference costly. To address these challenges, we propose PDScheduler, a neural scheduling framework built upon Persistent Dynamic Policy Decomposition (PDPD), which factorizes scheduling guidance into a persistent component and a dynamic residual. Specifically, PDScheduler employs masked discrete diffusion to generate a global assignment heatmap once per instance, while a lightweight residual actor dynamically adjusts the persistent guidance according to the evolving partial schedule. This design performs the heavy computation over the instance only once and shares the resulting global guidance across all subsequent decisions, eliminating the redundant recomputation that burdens existing constructive methods. Experiments on public FJSP benchmarks show that PDScheduler outperforms state of the art neural baselines in makespan under both in distribution and out of distribution evaluations while running 1.5× faster, demonstrating its superior solution quality and computational efficiency.

Primary area:
optimization
Keywords:
Flexible Job-Shop Scheduling, Combinatorial Optimization, Deep Reinforcement Learning, Diffusion Models, Policy Decomposition, Neural Scheduling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NThxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PDScheduler, a neural framework for flexible job-shop scheduling that decomposes action scores into a persistent instance-level prior and a dynamic residual. A masked discrete diffusion model produces an assignment heatmap and structural embeddings. These outputs are reused throughout schedule construction, while a lightweight residual actor adapts decisions to the evolving scheduling state.

Critical strengths

PDScheduler computes an assignment prior and structural embeddings once, then uses a lightweight residual actor to jointly select operations and machines as the schedule evolves. This avoids committing to machine assignments prematurely and reduces repeated global computation.

The method achieves the lowest mean gap in all eight distribution–size settings and improves aggregate performance when transferred to larger instances without retraining.

Critical weaknesses

The ablations do not clearly separate the benefits of diffusion training and policy decomposition from cached embeddings and teacher supervision. Proposition 1 also does not connect the learned heatmap to the ideal projection. This leaves the source of the performance gains unclear.

PDScheduler is faster but has worse aggregate solution quality than the strongest baselines on public FJSP benchmarks. Equal sampling counts do not ensure equal computational budgets. Without matched-time comparisons, the practical quality–efficiency advantage remains unclear.

Questions for authors

Is there a controlled comparison using the same teacher labels, cached encoder, and residual actor, but direct assignment prediction instead of diffusion training?

How do PDScheduler, RESCHED, and DANIEL compare under matched wall-clock budgets on public benchmarks? Does timing include the initial denoiser pass, and is evaluation batching comparable across methods?

6 comments

We thank the reviewer for the specific suggestions concerning controlled ablations, the interpretation of Proposition 1, and matched inference budgets. We have added comparisons that examine the training objective and the source of assignment guidance, clarified the structural ablations, and expanded the public-benchmark evaluation under matched wall-clock budgets. We jointly address each weakness and its corresponding question below.

W1Q1. Separating the effects of diffusion training, teacher supervision, and cached information

To separate these factors, we added controlled comparisons of the training objective and the source of assignment guidance. The Direct Supervised GNN uses the same GNN architecture, RESCHED assignment labels, caching interface, and residual actor architecture as PDScheduler. It replaces diffusion pretraining with standard cross-entropy for direct assignment prediction, and the resulting configuration is trained separately.

We also evaluated standard masked modeling using the same random masking procedure but without diffusion-specific loss weighting. These comparisons distinguish masked reconstruction from direct prediction, and diffusion-specific weighting from ordinary masked CE.

Results on 500 SD1 10×510\times5 instances. Lower mean Cmax⁡C_{\max} is better.

Guidance-learning configurationMean Cmax⁡C_{\max} ↓
Direct supervised GNN103.716
Standard masked CE101.488
Masked diffusion101.218

Standard masked CE reduces mean makespan from 103.716 to 101.488 compared with direct supervised prediction, and diffusion-specific weighting provides a further reduction to 101.218.

Direct teacher guidance. As detailed in Appendix A.5 (“RESCHED assignment guidance”), we replace the predicted heatmap scores with smoothed log-probabilities derived from RESCHED machine assignments, while retaining cached structural embeddings, the RL-trained residual actor, and the EFT bias. Other compatible machines remain selectable, allowing the actor to revise teacher assignments during schedule construction.

This variant obtains a mean makespan of 104.142, compared with 101.218 for PDScheduler. The comparison isolates the source of explicit assignment guidance while retaining learned structural embeddings, and is distinct from evaluating the complete RESCHED policy.

Updated structural ablations. We have corrected the structural-ablation results and clarified the interventions for the key variants:

VariantMean Cmax⁡C_{\max} ↓
Full model101.218
Static only133.384
Dynamic only134.094
w/o static embeddings126.206
w/o heatmap scores103.012
  • Static only: Disable the residual actor while retaining the explicit heatmap score and EFT bias. Cached embeddings do not participate in action scoring.
  • Dynamic only: Set both heatmap-score pathways and all cached static embeddings to zero, while retaining the actor and EFT.
  • w/o static embeddings: Remove the cached static embeddings supplied by the first-stage diffusion model to the actor. Retain both heatmap-score pathways, the residual actor, and the EFT bias.
  • w/o heatmap scores: Set heatmap scores to zero both within candidate representations and in their direct additive contribution to logits. Retain cached embeddings, the actor, and EFT.

These definitions distinguish information removal from disabling a network component. In particular, Static only retains a state-dependent EFT preference.

The results support the contribution of cached structural information and state-dependent scoring in the tested configuration.

Proposition 1. As clarified in our response to Reviewer fMVK, Proposition 1 characterizes the optimal state-independent approximation to prescribed logits under a fixed state distribution. The learned heatmap is trained from reference assignments and is not assumed to equal this ideal projection. The controlled comparisons above assess the empirical contributions of the training objective, the source of assignment guidance, and cached structural information. (continued in next comment)

W2Q2. Public-benchmark quality under matched wall-clock budgets and timing fairness

In the original Greedy and Sampling-100 evaluations, PDScheduler is faster, while RESCHED obtains lower aggregate gaps on the public benchmarks. Since equal sampling counts do not ensure equal computational budgets, we added matched wall-clock comparisons to evaluate solution quality under a common inference budget.

The experiment covers 130 public FJSP instances with five evaluation seeds per instance, giving 650 instance–evaluation-seed runs. We compare PDScheduler, RESCHED, and DANIEL at common checkpoints of 100 ms, 200 ms, 500 ms, 1 s, and 2 s.

For each checkpoint, only complete feasible schedules returned and incorporated into the candidate set before the deadline are counted. The best makespan among these schedules is retained.

Complete-schedule coverage and conditional mean Gap. Coverage is measured over all 650 runs. Mean Gap is calculated only over runs with a completed feasible schedule; “—” indicates that no run completes.

BudgetPDScheduler Gap (%)Coverage (%)RESCHED Gap (%)Coverage (%)DANIEL Gap (%)Coverage (%)
100 ms—0.00—0.00—0.00
200 ms11.11213.08—0.00—0.00
500 ms9.24463.089.63245.2310.34349.23
1 s8.579100.009.25867.699.37567.08
2 s7.336100.007.518100.007.862100.00

At 1 s, PDScheduler returns a complete feasible schedule in every run. Conditional mean gaps at budgets with different coverage are computed on different subsets and should not be interpreted as quality rankings on identical samples.

Comparison on common completed runs. We therefore additionally compare the methods on the subset of runs completed by all three methods at each checkpoint.

BudgetCommon completed runsPDScheduler Gap (%)RESCHED Gap (%)DANIEL Gap (%)
500 ms294/6506.7919.63210.256
1 s410/6506.1939.1979.212
2 s650/6507.3367.5187.862

Within each row, all methods are evaluated on identical runs. The subset changes across budgets, so values should not be compared across rows as a quality-versus-time trajectory.

PDScheduler obtains lower mean gaps than both baselines on the common completed subsets at 500 ms and 1 s. At 2 s, all methods complete all 650 runs. PDScheduler obtains a mean Gap of 7.336%, which is lower than RESCHED and DANIEL by 0.182 and 0.526 percentage points, respectively.

Timing and batching protocol. All methods were evaluated serially on the same NVIDIA GeForce RTX 5090 GPU. PyTorch intra-operation and inter-operation CPU thread counts were set to one, as were OMP, MKL, and OpenBLAS thread counts.

Each run evaluates one instance, first generating one greedy schedule and then sampling trajectories in batches of eight. Thus, the trajectory batch size and decoding order are matched across methods.

The time budget includes:

  • Per-instance environment initialization.
  • Input processing and data transfers.
  • PDScheduler’s initial persistent-denoiser forward pass.
  • Schedule decoding and best-candidate updates.

Persistent outputs are computed once per instance and reused across subsequent trajectories.

Model loading, test-file reading and parsing, and warmup are excluded consistently. All methods use the same warmup procedure: one greedy decode and one sampling batch on the first instance of each benchmark suite. Warmup results are discarded, and the instance state is reinitialized for formal evaluation.

Elapsed time is measured with time.perf_counter(). CUDA synchronization is performed before timing begins, after instance preparation, and after each decoding batch. Each run records results at the five checkpoints up to 2 s.

Only schedules fully returned and incorporated into the candidate set before a checkpoint are included. If a sampling batch crosses a deadline, its results are excluded from that checkpoint. Runs without a completed schedule before the deadline are recorded as having no feasible result.

Under the matched-time protocol, PDScheduler achieves full coverage at 1 s and the lowest aggregate mean Gap among the three methods at 2 s, when all methods achieve full coverage.

Thank you for providing the additional experiments which addresse my previous concerns. But I have one remaining concern: the robustness of the performance gains attributed to "diffusion training." In Table 2, the full model achieves an average makespan of 101.218, compared to 101.488 for the standard masked CE. This difference is relatively small, and the comparison is limited to a single problem scale "SD1 10×5." Could the authors provide more details regarding the distribution of these performance improvements? Furthermore, extending the comparison to other problem scales would help verify whether these gains generalize beyond the current experimental setup.

Should further supplementary analysis support this, I would be open to reconsidering my score.

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

This paper proposes PDScheduler for FJSP: decompose policy guidance into a persistent prior and a dynamic residual. A masked discrete diffusion denoiser runs once per instance and produces a soft machine-assignment heatmap plus structural embeddings; a lightweight residual actor then reweights this prior over feasible operation–machine pairs using the evolving partial schedule. Proposition 1 states that the L2-optimal fixed score is the conditional mean of prescribed logits. Experiments show better synthetic FJSP makespans than HGNN/DANIEL/RESCHED, competitive-but-not-best results on public FJSP suites at ~1.8× lower inference cost, and secondary JSSP/FFSP variants.

Critical strengths

Clear design tension (assignment-first vs. fully autoregressive) and a practical resolution: global guidance once, state-dependent correction every step, without fixing machines. Solid experimental hygiene: shared instances, multi-seed runs, ablations (static-only / dynamic-only / no heatmap), cross-size transfer, and honest reporting of public-benchmark gaps. Inference is genuinely faster than the neural baselines, with quality largely preserved on synthetic and larger instances.

Critical weaknesses

The theory is essentially vacuous. Proposition 1 is the elementary fact that the L2 projection of a random vector onto constants is its mean; the paper itself admits it neither bounds actor capacity nor identifies the learned heatmap with that projection. It does not connect to makespan or justify the architecture. Not SOTA on public benchmarks. RESCHED/DANIEL are stronger on Brandimarte, rdata, and vdata; the 130-instance aggregate is worse than RESCHED (11.60% vs 10.74% greedy; 5.93% vs 5.28% Sampling-100). The real claim is “competitive + faster,” which the abstract oversells as outperforming SOTA. The diffusion prior is weakly motivated. Default is a single forward pass from fully masked input (K=1), and refinement does not help. Ablation shows a plain “Direct GNN” is only ~2.5% worse — so the expensive masked-diffusion machinery is not clearly necessary versus a supervised heatmap network. Denoiser labels are also distilled from RESCHED’s greedy assignments, so independence from the teacher is limited (end-to-end training is left as future work). “Prior + residual” is a familiar pattern (DIFUSCO/T2T heatmaps + decoder, COExpander, MatNet reuse). JSSP/FFSP results are mixed and often trail RESCHED on quality; no comparison to diffusion-CO pipelines or local-search hybrids that are standard for strong FJSP solvers.

Questions for authors

Why a masked diffusion denoiser at all under K=1? Please compare to a capacity-matched supervised heatmap GNN (and to RESCHED-only guidance) on the same splits. How does PDScheduler compare to DIFUSCO/T2T/COExpander-style pipelines or LS hybrids on the same FJSP benchmarks? Can you train without RESCHED-derived assignment labels (or end-to-end) and report whether the synthetic gains remain? On Brandimarte/rdata/vdata where you trail RESCHED/DANIEL, what failure mode under distribution shift causes the gap?

6 comments

We thank the reviewer for the detailed assessment of the theoretical motivation, diffusion pretraining, teacher supervision, and experimental scope. We have revised the contribution statements and added controlled experiments to distinguish the roles of guidance learning, cached information, and state-dependent scheduling. Below, we jointly address overlapping weaknesses and questions.

W1. Proposition 1 and the architectural motivation

Proposition 1 applies the standard squared-error projection identity to characterize the optimal state-independent approximation to prescribed logits under a fixed state distribution. In PDScheduler, the heatmap is learned from reference assignments and is not assumed to equal this ideal projection.

The architectural motivation comes from the distinction between invariant and evolving information in FJSP. Processing times, machine eligibility, and precedence constraints remain fixed within an instance, while machine availability, job readiness, and competition among candidate actions change during schedule construction. This motivates computing reusable instance representations once and adapting their contribution to subsequent decisions.

The practical contributions are evaluated empirically: pretraining comparisons examine how guidance is learned, structural ablations examine information inputs and actor use, and matched-time experiments evaluate the quality–efficiency trade-off of the complete system. (continued in next comment)

W2Q4. Public-benchmark performance, matched inference budgets, and distribution shift

We have revised the abstract and experimental discussion to distinguish synthetic performance from public-benchmark results. The original Greedy and Sampling-100 results are:

DecodingPDScheduler Gap (%)RESCHED Gap (%)DANIEL Gap (%)
Greedy11.6010.7810.74
Sampling-1005.935.285.41

For greedy decoding, PDScheduler reduces the average inference time from 0.681 s for RESCHED to 0.378 s per instance, corresponding to approximately 1.8×1.8\times faster inference. These fixed-count results show a trade-off between solution quality and inference time.

To examine whether the inference advantage translates into better results under a common computational budget, we additionally compared the methods under matched wall-clock budgets on 130 public instances with five evaluation seeds per instance, giving 650 instance–evaluation-seed runs.

At 1 s, PDScheduler returns a complete feasible schedule in every run, whereas RESCHED and DANIEL achieve coverage of 67.69% and 67.08%, respectively. At 2 s, all methods achieve full coverage, allowing comparison on the same 650 runs.

Matched 2-second comparison on public FJSP benchmarks. Lower mean Gap is better.

BenchmarkInstancesPDScheduler Gap (%)RESCHED Gap (%)DANIEL Gap (%)
Brandimarte109.6458.54410.626
Hurink edata409.07611.41913.145
Hurink rdata407.5238.0758.346
Hurink vdata404.8312.8061.405
Instance-weighted average1307.3367.5187.862

At the 2 s budget, PDScheduler obtains the lowest aggregate mean Gap among the three methods, improving on RESCHED and DANIEL by 0.182 and 0.526 percentage points, respectively. It also obtains the lowest mean gaps on edata and rdata; RESCHED obtains the lowest mean Gap on Brandimarte, and DANIEL on vdata.

All methods were evaluated serially on the same RTX 5090 GPU with controlled CPU thread counts. Timing includes instance initialization, input processing and transfers, the initial denoiser computation for PDScheduler, decoding, and candidate updates. Model loading, test-file parsing, and warmup are excluded consistently. Only complete feasible schedules returned and incorporated into the candidate set before each deadline are counted. The complete checkpoint results and timing protocol are provided in our response to Reviewer NThx, W2Q2.

Regarding distribution shift, the ranking change on rdata shows that the comparison is sensitive to the inference budget. At 2 s, the remaining solution-quality gaps occur on Brandimarte and vdata. The current experiments identify these performance differences but do not isolate a specific distribution-shift mechanism. (continued in next comment)

W3Q1. Why diffusion under K=1K=1? Controlled supervised and teacher-guidance comparisons

At K=1K=1, the pretrained predictor makes one fully masked prediction per instance to produce assignment guidance and cached structural embeddings. To evaluate the pretraining objective under this inference setting, we compared direct supervised prediction, standard masked modeling, and masked diffusion using the same GNN architecture, RESCHED assignment labels, caching interface, and residual actor architecture. Each variant is trained separately.

We also added a comparison that directly uses RESCHED assignments as guidance while retaining structural embeddings and an RL-trained residual actor.

Results on 500 SD1 10×510\times5 instances. Lower mean Cmax⁡C_{\max} is better.

ConfigurationMean Cmax⁡C_{\max} ↓
Direct supervised GNN103.716
Standard masked CE101.488
RESCHED assignment guidance + embeddings + RL actor104.142
PDScheduler101.218

Standard masked CE reduces mean makespan from 103.716 to 101.488 compared with direct supervised prediction, and diffusion-specific weighting provides a further reduction to 101.218. These comparisons show that PDPD accommodates multiple guidance-learning objectives, with masked diffusion obtaining the lowest mean makespan in this comparison.

Direct RESCHED guidance. As detailed in Appendix A.5 (Experimental details, “RESCHED assignment guidance”), we replace the predicted heatmap scores with smoothed log-probabilities derived from RESCHED machine assignments. The residual actor is trained under this guidance while retaining cached structural embeddings and the EFT bias. Other compatible machines remain selectable, allowing the actor to revise teacher assignments during schedule construction. This variant obtains a mean makespan of 104.142, compared with 101.218 for PDScheduler, supporting the benefit of learned assignment scores over direct teacher guidance in this configuration.

We have also updated the refinement analysis. Across the eight SD1 and SD2 settings, K=1K=1 obtains the lowest mean makespan in five settings. Increasing KK from 1 to 8 raises the average inference time from 38.56 ms to 50.05 ms per instance, while the largest relative makespan reduction over K=1K=1 among the tested alternatives is approximately 0.324%. These results support K=1K=1 as a practical default for balancing solution quality and inference cost, with small gains from additional refinement in some settings.

The refinement experiment evaluates additional inference-time computation, while the controlled pretraining comparisons evaluate the learning objective.

W4Q3. Dependence on RESCHED labels and end-to-end training

In the proposed two-stage procedure, RESCHED machine assignments supervise only denoiser pretraining. The actor is subsequently optimized using the makespan of completed schedules and does not imitate the teacher’s action sequence.

To examine training without assignment supervision, we jointly trained the denoiser and residual actor from scratch using REINFORCE, preserving their architectures and using no RESCHED assignment labels.

Training strategyMean Cmax⁡C_{\max} ↓
Two-stage training101.218
End-to-end joint REINFORCE101.548

On SD1 10×510\times5, joint REINFORCE learns a policy without RESCHED assignment labels, obtaining a mean makespan of 101.548 compared with 101.218 for two-stage training.

This comparison covers SD1 10×510\times5 and evaluates solution quality without matching total training cost, which additionally includes label generation and denoiser pretraining for the two-stage procedure.

Both training approaches retain the ability to compute and cache instance-level outputs during inference. (continued in next comment)

Reviewer_9LXdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the Flexible Job-Shop Scheduling Problem (FJSP) using a GNN-based learning-to-construct framework that sequentially selects operation–machine (O–M) pairs. The main idea is to separately process static and dynamic features using two independently trained networks. The static network is pretrained using a diffusion-based learning objective and then frozen during policy training and inference, while the dynamic network is optimized using reinforcement learning. Experimental results show promising performance on synthetic FJSP instances, particularly on SD1. However, the advantages are less consistent on SD2, public FJSP benchmarks, and the JSSP and FFSP evaluations.

The technical motivation is unconvincing, and the presentation is not easy to follow. This paper introduces the concepts of "Persistent Score" and "Dynamic Residual" to describe the separation of static and dynamic information/features, while using a relatively complicated training method that combines diffusion-based pretraining with RL training. In my opinion, these design choices are insufficiently motivated, and their necessity is not convincingly demonstrated. At this stage, I would give it a 1.5 rating, and may adjust this depending on the rebuttal.

Critical strengths
  1. The paper adopts standard and widely used scheduling benchmarks for experimental evaluation.
Critical weaknesses
  1. The proposed framework combines several existing techniques without adequately explaining why they are necessary. In particular, diffusion learning is presented as an important component, but its practical necessity has not been sufficiently demonstrated. Similarly, the introduction of an explicit earliest finish time (EFT) bias appears ad hoc and lacks clear justification. Overall, the paper does not provide sufficient evidence that the proposed combination of techniques is necessary or preferable to simpler alternatives.

  2. Several concepts are not clearly distinguished, and the presentation lacks a coherent logical flow. For example, the categorization in Figure 1 is misleading, since both (b) and (c) are essentially learning-to-construct approaches. The author incorrectly conflates the concepts of "autoregressive" and "constructive." In addition, the main text does not sufficiently explain the network architecture and the specific neural components used in the proposed framework.

Questions for authors
  1. Since the denoiser performs only a single fully masked prediction during inference, why is diffusion training necessary instead of directly training a supervised GNN? Please compare against: (1) Use the same GNN architecture using supervised cross-entropy. (2) Use the same GNN architecture with standard random masked modeling instead of diffusion-specific loss weighting.

  2. The explicit EFT-based bias in the residual action scores appears somewhat ad hoc. Why is EFT introduced as an additional heuristic preference? Could other dispatching heuristics be used instead?

  3. Could the authors compare the proposed two-stage training procedure with an end-to-end alternative that preserves the same static and dynamic network architectures but jointly optimizes both components using REINFORCE?

  4. As shown in Table 4, increasing the persistent refinement budget KK provides little or no improvement in scheduling quality. In other words, this indicates that PDScheduler does not derive any significant benefit from Iterative Diffusion Refinement. Could the authors explain the raised questions about the practical benefits of iterative diffusion refinement?

  5. The OR-Tools reference results for SD2 in Table 1 appear questionable, which raises concerns about the reliability of the corresponding gap values.

  6. Although PDScheduler performs well on SD1, its advantages are less consistent on SD2, public FJSP benchmarks, JSSP, and FFSP. Despite introducing a more complicated architecture and training procedure, it does not consistently outperform RESCHED. In addition, the experiments report only three training seeds, fewer than the commonly used five seeds.

  7. Several design choices require further clarification:

    • Why is RESCHED used to generate the machine assignment samples instead of a stronger solver such as OR-Tools?
    • What is the justification for the chosen masking probability function? Why is this particular schedule adopted?
    • The recurrent memory ct\mathbf{c}_t is not clearly explained. What information does it contain, and how does it contribute to scheduling decisions?
    • Could the authors provide the training curves for the 60 epochs of denoiser pretraining?
    • Why are the bounds for the implemented cost advantage set to (-3, 3)? What is the rationale behind this clipping?
9 comments

We thank the reviewer for the detailed comments on the technical motivation, presentation, and experimental evaluation. We have clarified the roles of the individual components, added controlled comparisons with simpler alternatives, corrected the SD2 experiments, and expanded the evaluation of inference efficiency and training variability. The original manuscript has also been updated, with changes in blue. Below, we address each concern.

W1Q1. Technical motivation and the role of diffusion pretraining

FJSP contains fixed structural information, including processing times, machine eligibility, and precedence constraints, while the suitability of operation–machine pairs changes with the evolving schedule. PDPD addresses this distinction by computing reusable assignment guidance and structural embeddings once, while a residual actor adjusts action scores based on the current scheduling state, jointly determining machine assignments and operation sequencing.

Diffusion pretraining is motivated by the presence of multiple competitive assignment patterns in FJSP. With only one reference assignment per instance, direct cross-entropy supervision penalizes alternative choices that may also yield competitive schedules. We therefore learn soft assignment guidance through masked reconstruction across different corruption levels. By conditioning on partially observed assignments and graph structure, the denoiser learns dependencies among machine choices, while diffusion-specific weighting further adjusts the reconstruction objective.

At inference, a single fully masked prediction produces cached guidance, while the residual actor explores scheduling alternatives through sampled trajectories. To evaluate the training objective, we compare direct supervised GNN, standard masked CE, and masked diffusion using the same GNN architecture, RESCHED labels, caching interface, and residual actor, with each configuration trained separately. These comparisons isolate the effects of direct prediction, masked reconstruction, and diffusion-specific weighting.

Updated ablation results on 500 SD1 10×510\times5 instances. Lower mean Cmax⁡C_{\max} is better.

VariantMean Cmax⁡C_{\max} ↓
Full model101.218
Direct supervised GNN103.716
Standard masked CE101.488
Static only133.384
Dynamic only134.094
w/o static embeddings126.206
w/o heatmap scores103.012
End-to-end joint REINFORCE101.548
Without EFT bias101.912
SPT bias instead of EFT101.734

Standard masked CE reduces mean makespan from 103.716 to 101.488 compared with direct supervised prediction, and diffusion-specific weighting provides a further reduction to 101.218. These comparisons show that PDPD accommodates multiple guidance-learning objectives, with masked diffusion obtaining the lowest mean makespan in this comparison.

We have also clarified the operational definitions of the structural ablations:

  • Static only: Disable the residual actor while retaining the explicit heatmap score and EFT bias. Cached embeddings do not participate in action scoring.
  • Dynamic only: Set both heatmap-score pathways and all cached static embeddings to zero, while retaining the residual actor and EFT bias.
  • w/o static embeddings: Remove the cached static embeddings supplied by the first-stage diffusion model to the actor. Retain both heatmap-score pathways, the residual actor, and the EFT bias.
  • w/o heatmap scores: Set heatmap scores to zero both within candidate representations and in their direct contribution to action logits. Retain cached embeddings, the actor, and EFT.

These definitions distinguish information removal from disabling a network component. In particular, Static only retains a state-dependent EFT preference. The results support the contribution of cached structural information and state-dependent scoring in the tested configuration. (continued in next comment)

Q2. Motivation for EFT and alternative dispatching heuristics

EFT is an established dispatching heuristic that favors earlier operation completion. We incorporate it as an explicit timing preference because it accounts for both processing time and the earliest feasible start time determined by job readiness and machine availability under the current partial schedule.

As defined in Section 4.2 of the revised manuscript, the action score combines the persistent assignment score, the learned state-dependent residual, and an instance-normalized EFT bias. The EFT term provides an explicit timing preference based on job readiness and machine availability, while the learned residual can adjust the action ranking to optimize the final makespan rather than follow a fixed dispatching rule.

The EFT term provides a timing preference rather than enforcing a fixed dispatching rule. The learned residual can adjust the action ranking to optimize the final makespan. The choice of heuristic is not a methodological contribution of PDScheduler; instead, the heuristic term serves as an interchangeable preference within the action-scoring function.

To examine its contribution and flexibility, we compared EFT with another established dispatching heuristic, Shortest Processing Time (SPT), and a variant without an explicit heuristic bias.

Explicit scoring biasMean Cmax⁡C_{\max} ↓
EFT101.218
SPT101.734
None101.912

Both EFT and SPT reduce mean makespan relative to removing the explicit bias, with EFT obtaining the lowest mean on SD1 10×510\times5. These results demonstrate that the scoring function accommodates alternative heuristic preferences, while EFT provides a stronger timing preference in this setting.

We have clarified the motivation for EFT and its role as an interchangeable heuristic preference in Section 4.2.

Q3. Two-stage training versus end-to-end REINFORCE

The two training stages serve different objectives. Denoiser pretraining learns machine assignment guidance from reference schedules. The subsequent RL stage freezes the denoiser and optimizes the actor using the makespan of completed schedules. The actor does not imitate the teacher’s action sequence.

To compare this procedure with joint optimization, we evaluated an alternative that preserves the denoiser and residual actor architectures but jointly trains both components from scratch using REINFORCE, without RESCHED assignment supervision.

Training strategyMean Cmax⁡C_{\max} ↓
Two-stage training101.218
End-to-end joint REINFORCE101.548

On SD1 10×510\times5, joint REINFORCE obtains a mean makespan of 101.548 without RESCHED assignment labels, compared with 101.218 for two-stage training. Both approaches retain the ability to compute and cache instance-level outputs during inference.

This comparison evaluates solution quality without matching total training cost, which additionally includes label generation and denoiser pretraining for the two-stage procedure. (continued in next comment)

W2. Clarity of presentation, conceptual distinctions, and network architecture

PDScheduler constructs schedules autoregressively by sequentially selecting feasible operation–machine pairs. We have revised Figure 1 and the accompanying discussion to clarify that both approaches are autoregressive and constructive. Their distinction concerns how representations are computed and reused.

We have made the following revisions:

  1. Figure 1 and Introduction: We renamed Figure 1(b) from “Autoregressive” to “Repeated State Encoding” and clarified that both Figure 1(b) and PDScheduler are autoregressive constructive approaches. Their distinction lies in whether full-graph representations are repeatedly computed or instance-level guidance is reused.

  2. Method overview and Figure 2: We clarified the complete information flow, distinguishing the persistent outputs computed once per instance from the state-dependent corrections updated at each decision step. The revised description explains how the two components jointly score feasible operation–machine pairs.

  3. Section 4.1 (Persistent Denoiser): We clarified the denoiser inputs, heterogeneous graph message-passing architecture, and outputs, including assignment probabilities and operation, machine, edge, and graph-level embeddings. We also explicitly distinguished masked diffusion pretraining from the single fully masked evaluation used at inference.

  4. Section 4.2 (Residual Actor): We expanded the description of candidate representations, self-attention over feasible operation–machine pairs, and GRU memory updates. We also made the EFT-based timing preference explicit in the final action-scoring equation.

  5. Training procedure: We added a concise description of the two training stages, specifying their optimization objectives, parameter updates, and the caching mechanism used during policy training and inference.

These revisions clarify the roles and interactions of the components, particularly the distinction between reusable instance information and the evolving scheduling state. (continued in next comment)

Comments on the paper

1 comment

Dear Reviewers,

We would like to summarize the main experiments and revisions made in response to your comments. All changes are marked in blue in the revised manuscript.

  • Pretraining objective and teacher guidance (NThx, fMVK, 9LXd): Added controlled comparisons with direct supervised GNN prediction and standard masked cross-entropy using the same architecture, assignment labels, and actor interface. We also evaluated direct RESCHED assignment guidance while retaining cached structural embeddings and an RL-trained actor.

  • Diffusion comparison across scales (NThx): Extended the masked-CE comparison to three additional SD1 scales: 20×5, 15×10, and 20×10. The full model obtains lower mean makespans across all four tested SD1 scales, extending the original 10×5 comparison.

  • Structural ablations and joint training (NThx, 9LXd): Clarified the interventions in the static-only, dynamic-only, embedding-removal, and heatmap-removal variants. We also added an alternative that jointly trains the same static and dynamic network architectures using REINFORCE on SD1 10×5.

  • Heuristic bias and refinement budget (9LXd, fMVK): Compared EFT with SPT and no explicit bias, demonstrating the replaceability of the heuristic component. We updated the refinement analysis and clarified the distinction between diffusion pretraining and iterative inference; the quality–time results support K=1 as a practical default.

  • SD2 evaluation and reference validation (9LXd): Regenerated SD2 using the processing-time range specified in the manuscript, fully retrained both stages of PDScheduler, and re-evaluated all methods on the same test instances. We recomputed the OR-Tools references and documented the solver settings, feasibility and optimality status, and gap definitions.

  • Matched wall-clock evaluation (NThx, fMVK, 9LXd): Added comparisons on 130 public benchmark instances with five evaluation seeds per instance, reporting complete-schedule coverage and quality on common completed runs. At 2 seconds, all methods achieve full coverage, with mean gaps of 7.336% for PDScheduler, 7.518% for RESCHED, and 7.862% for DANIEL.

  • Training variability and learning curves (9LXd): Added five independent, complete training runs on SD1 20×10, including both denoiser pretraining and actor training. We also provided training and validation curves covering the full 60-epoch denoiser pretraining procedure.

  • Technical motivation and presentation (NThx, fMVK, 9LXd): Clarified the scope of Proposition 1, the roles of cached structural information and state-dependent scoring, and the network components. We also revised the terminology in Figure 1 to distinguish computational reuse from the autoregressive construction of schedules.

We thank the reviewers for their time and constructive suggestions. We remain available to clarify any remaining questions about the additional experiments or the revised manuscript.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.