acceptodds
Under review as a conference paper at ICLR 2027

CausaSched: A Causal-Guided Deep Reinforcement Learning Hyper-Heuristic for Improving Multiple Shop Scheduling Problems

Abstract

Traditional constructive heuristics and most learning-based scheduling methods generate schedules from scratch, whereas recent learning-based improvement methods start from existing feasible solutions and further improve their quality through successive neighborhood moves. Because schedule improvement concerns how changes in controllable decisions affect the makespan, this process can be modeled from a causal-intervention perspective. Based on this observation, we introduce causal reasoning into learning-based hyper-heuristics and propose CausaSched for multiple scheduling problems. Causal reasoning starts from scheduling bottlenecks, traces how delays arise through structural dependencies, and determines where subsequent schedule modifications should be applied. Achieving this goal raises three key questions: how to efficiently search for intervention targets that can reduce makespan, how to match them with neighborhood operators satisfying problem-specific constraints, and how to jointly train both decision networks to improve search performance. CausaSched exploits job-precedence and machine-order dependencies shared across scheduling problems: a causal reasoning network learns where to intervene, and an operator selection network determines the neighborhood move. We further propose Causal Hierarchical Group Relative Policy Optimization (CH-GRPO), which jointly trains both networks using multiple multi-step search trajectories and expands the neighborhoods of high-quality states. Experiments on public benchmark instances and mixed synthetic instances show that CausaSched significantly outperforms deep reinforcement learning and improvement heuristics.

Primary area:
optimization
Keywords:
shop scheduling, causal reasoning, deep reinforcement learning, hyper-heuristics, graph neural networks, combinatorial optimization

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_ubJcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CausaSched, a causal-guided deep reinforcement learning hyper-heuristic framework for multiple shop scheduling problems. The authors suggest identifying the sources of delays in the current schedule and then using an operation selection network to make adjustments. The experiments were conducted on several public scheduling benchmarks and mixed synthetic datasets, comparing it with DRL constructive methods and traditional boosting methods. The results show that CausaSched can achieve better makespan optimization in most test scenarios.

Critical strengths
  1. The motivation is well-grounded, and the idea of locating intervention targets via delay propagation provides a fresh perspective for learning-based scheduling optimization.
  2. The cross-problem generalization design is systematic, and evaluations on multiple public benchmarks verify its empirical effectiveness.
Critical weaknesses
  1. The so-called "causal reasoning" is essentially learned weight propagation along scheduling dependencies, not formal causal inference. The paper provides neither strict counterfactual effect estimation nor empirical proof that LCT-identified targets are truly the root causes of makespan delay.
  2. The paper compares an iterative improvement method against one-pass constructive DRL baselines, which have completely different computational paradigms and inference budgets. Meanwhile, it lacks full head-to-head comparison with state-of-the-art DRL improvement methods (e.g., MIStar for FJSP).
  3. Employing multi-trajectory search and candidate intervention mechanisms during inference may incur additional computational overhead.
  4. A scheduling graph does not necessarily represent true causal relationships; instead, it may primarily reflect feasibility constraints. Consequently, what the LCT learns is likely correlation within the graph structure rather than genuine causal influence.
  5. All ablations are conducted at inference time on a fixed checkpoint, without retraining ablated variants. This only proves component importance for the current trained weights, not their necessity during learning.
Questions for authors

See Weaknesses.

Reviewer_9QCZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CausaSched, a reinforcement learning hyper-heuristic for improving feasible schedules across JSP, FSP, FJSP, and HFSP. The framework combines learned upstream tracing for intervention target selection, single and compositional neighborhood operators, and an intervention memory. The authors introduce CH-GRPO to jointly train target and operator selection using step-level and trajectory-level feedback, and report promising results on public benchmarks and mixed synthetic instances.

Critical strengths

S1. The framework supports four scheduling families through shared dependency representations and constraint-aware operators, offering a useful approach to unified schedule improvement.

S2. Compositional operators can capture complementary modifications whose joint benefit is overlooked when individual moves are evaluated independently.

S3. Experiments across multiple public benchmarks and mixed synthetic instances demonstrate promising solution improvements over the evaluated constructive DRL methods and rule-based improvement heuristics.

Critical weaknesses

W1. The distinctive contribution of causal reasoning is insufficiently established relative to learned dependency tracing and conventional neighborhood evaluation.

W2. The evaluation omits closely related learning-based improvement methods such as MIStar, limiting the evidence for the framework’s comparative advantage.

W3. The contribution of CH-GRPO is not isolated through comparisons with standard policy optimization methods or ablations of its step-level and trajectory-level feedback.

Questions for authors

Q1. Could the authors clarify what distinguishes LCT’s learned responsibility propagation from reverse message passing over scheduling dependencies, and its counterfactual evaluation from standard neighborhood replay? The Uniform-LCT comparison demonstrates the usefulness of learned propagation weights, but provides limited evidence for the specific benefit of causal reasoning. A matched learned dependency-scoring baseline, or an analysis relating tracing scores to actual intervention gains, would help establish this contribution.

Q2. Could the authors explain the omission of MIStar, which is discussed as a closely related FJSP improvement method, and provide a direct comparison? Since the main FJSP learning baselines are constructive methods, comparison with a competitive learned improvement approach is important for assessing the claimed advances. Matching initial schedules and reporting solution quality under comparable runtime budgets would make this comparison more informative.

Q3. Could the authors compare CH-GRPO with PPO or standard GRPO while keeping the architecture, initialization, and search procedure fixed? Please also evaluate variants using only step-level feedback and only trajectory-level feedback. These controls would clarify whether the proposed training objective contributes beyond the network architecture and multi-trajectory search.

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

This paper proposes CausaSched, a DRL-based method that improves existing schedules by selecting which operations to change and how to change them. The method is tested on four shop scheduling problems and achieves better results than several existing methods.

Critical strengths
  1. The method can handle four different scheduling problems using a single model.

  2. The authors test their method on several public datasets and provide ablation studies to show how different parts of the method affect its performance.

Critical weaknesses
  1. The authors argue that schedule improvement can be modeled from a causal-intervention perspective, which motivates the proposed CausaSched framework. However, the motivation here is not very convincing. In DRL methods, the Markov Decision Process (MDP) itself already captures how actions affect subsequent states and rewards, which seems to reflect the causal-intervention perspective. The so-called causal reasoning here is basically the function of an MDP in DRL. It does not seem new, just another name.

  2. The authors claim that constructive DRL methods cannot revise earlier decisions, while improvement learning offers a more flexible way to further reduce the makespan. Are there any references or evidence to support this claim? I don't think this necessarily holds, especially in dynamic scenarios.

  3. The authors state that "Critical-path-guided methods can discover effective modifications by repeatedly applying neighborhood search." Any references to support this claim?

  4. The authors claim that CausaSched can address multiple scheduling families, including JSP, FSP, FJSP, and HFSP, by exploiting shared scheduling dependencies and constraint-aware operations. Why can CausaSched be so generalizable across these different scheduling problems? The authors should explain why these shared dependencies are sufficient to support a unified scheduling policy.

  5. The authors claim that CausaSched outperforms constructive DRL methods and improvement heuristics. However, CausaSched itself is a DRL-based improvement method. Why not compare with other DRL-based improvement methods? Table 2 mainly compares against constructive DRL methods (DANIEL and HGNN) and traditional improvement heuristics (GD, FI, and BI). Although L2S is included for JSP in Table 3, comparisons with DRL-based improvement methods for FJSP are still missing.

  6. Figures 3 and 4 look very much as if they were generated automatically. The arrows are weird; some of them overlap and make the figures difficult to follow. A human-drawn figure or diagram should not be in this style.

Questions for authors

The paper's motivation or novelty does not convince me, and the writing is too abstract and hard to follow, with some figures that look AI-generated. I also doubt the large improvements over existing baselines, especially since the authors did not compare with enough strong DRL-based methods or provide code to check their results. Overall, I have serious concerns about this paper and recommend Reject.

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

The paper proposes CausaSched, a causal-guided hyper-heuristic framework for improving complete schedules across JSP, FSP, FJSP, and HFSP. Its first contribution is a learned causal tracing policy that follows delay propagation backward along job and machine dependencies to identify promising intervention targets. Its second contribution is a constraint-aware operator selection policy that chooses single or compositional neighborhood moves, such as rerouting, swapping, and insertion, while respecting family-specific feasibility constraints. The two policies are trained jointly with CH-GRPO, which combines step-level and trajectory-level feedback. Experiments on public benchmarks and mixed synthetic instances show that CausaSched generally outperforms constructive deep reinforcement learning and improvement heuristics in most size groups. Ablation results show that learned tracing, compositional operators, intervention memory, cross-attention, and graph attention each contribute to performance. The main conclusion is that analyzing delay sources before choosing an operator can make learned schedule improvement more effective and more transferable across scheduling families.

Critical strengths
  1. The paper does not restrict itself to a single scheduling family. It covers JSP, FSP, FJSP, and HFSP, trains on 128 instances across all four families, and evaluates on multiple public benchmarks plus a mixed synthetic set.
  2. Involving causality into the learning-based improvement heuristic is interesting.
  3. The paper is well-organized.
Critical weaknesses
  1. The paper does not directly validate that the traced targets are the actual delay sources. Uniform-LCT only shows that non-uniform propagation beats uniform propagation. It does not show that high-tracing-score targets yield higher intervention success rates, nor does it report the correlation between tracing score and realized gain, nor whether traced paths actually lie on critical delay chains. Without path-level diagnosis, causal tracing remains a useful exploration prior rather than a verified root-cause localization mechanism.
  2. Strong baselines are missing. The ablations include Uniform-LCT but not random target selection with the same operator pool, critical-path target selection with the same operators, or local-feature target selection with the same operators. Without these controls, it is hard to tell whether gains come from causal tracing or simply from greater network capacity and better exploration. Notably, removing cross-attention and GAT causes the largest drops, which suggests performance may rely heavily on state encoding rather than causal tracing itself.
  3. The ablations are inference-time ablations on a shared checkpoint, not ablations after retraining. The paper acknowledges this, but it means removing a component does not let the network adapt, so the reported drops may overstate or understate each component's role.
  4. There are no multiple seeds, confidence intervals, or significance tests.
  5. The figures are all small and hard to see.
  6. The causality discussed in this paper is different from the one we are familiar with. Using this term may not be appropriate.
Questions for authors
  1. The JSSP benchmark is not exghausted. You only use Taillard. Is this cherry-picked?
  2. Please clarify whether hardware and implementation differences are controlled enough for the reported timings to be comparable.
  3. On HFSP, the w/o GAT variant slightly outperforms the full model. How do the authors explain this result? Does it suggest that GAT is not reliably beneficial for HFSP? The authors should report results across multiple random seeds and different candidate counts to clarify whether this reversal is stable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.