Lifelong Learning of Scheduling Rules with Surrogate-assisted Genetic Programming for Dynamic Flexible Job Shop Scheduling

Luyao Zhu, Fangfang Zhang, Yi Mei, Ke Chen, Mengjie Zhang

Oct 30, 2026arXiv:2610.41158v1

cs.AI

Scorecard· 16/16

5.5/10 impact

Competent, well-evaluated first treatment of forgetting in GP for DFJSS, but incremental over prior surrogate-assisted GP work, small task sequence, and niche audience.

Abstract

Genetic programming (GP) has been widely used to learn scheduling rules for dynamic flexible job shop scheduling (DFJSS). However, trained rules tend to only generalise to tasks of similar scales and distributions, and when task characteristics change, learned rules often suffer significant performance degradation. Moreover, fine-tuning pre-trained rules on new tasks often leads to catastrophic forgetting, where prior knowledge is lost. This motivates the need for GP methods that can adapt to new tasks while preserving past knowledge, leading to a lifelong learning paradigm. However, lifelong learning in GP for DFJSS remains largely under-explored. To address this gap, this paper proposes a surrogate-assisted GP method for lifelong learning to handle sequential heterogeneous DFJSS tasks with job or batch arrivals. The framework includes three key components: an adaptive two-stage training strategy to balance plasticity and stability, a task-similarity-based weighting mechanism to regulate knowledge reuse, and an improved surrogate model with distance constraints to enhance estimation accuracy. Experimental results show that the proposed method effectively mitigates catastrophic forgetting while maintaining strong adaptability, outperforming several compared methods and demonstrating the potential of lifelong GP in dynamic scheduling environments.

AI Impact Assessments

(1 model)

Core Contribution

The paper proposes LLSGP, a lifelong-learning framework for genetic programming (GP) that evolves scheduling rules for dynamic flexible job shop scheduling (DFJSS) across a sequence of heterogeneous tasks. It has four ingredients:

Elite transfer: elites from the previous task seed the next task's population.

Adaptive two-stage training: Stage 1 runs standard GP on the current task. When the mean normalised population fitness stagnates, Stage 2 preselects offspring from a 3× oversized intermediate population using a combined fitness. That fitness mixes real current-task fitness with surrogate-estimated fitness on all earlier tasks, which acts as experience replay.

QD-based KNN surrogate with a distance threshold: samples are collected throughout training, kept only if they beat a reference rule, and are diverse in behaviour space. Candidates farther than τ from every sample get infinite fitness and are discarded.

Task-similarity weighting: Spearman correlation of fitness across tasks sets the weights, with less similar tasks weighted more.

The work is among the first to treat catastrophic forgetting explicitly in GP-based hyper-heuristics. Prior GP lifelong work (Kattan, Sim, Hart) was essentially fine-tuning or heuristic repertoires. The framing is useful, and the engineering is sensible, because surrogates make replay affordable for a population-based symbolic learner.

Methodological Rigor

Setup: four tasks (E/G/U/N processing-time distributions, differing scale, utilisation and batch settings), six training orders, and 30 runs each. Metrics (AP, AF, AMF, APl) follow continual-learning conventions and are normalised to a joint multi-task upper bound. Wilcoxon and Friedman tests are used.

Baselines: Restart, three fine-tuning ratios, LLR-BC (adapted from neural VRP, so only an idea-level adaptation), and a multi-task reference. The baselines are reasonable but not strong. There is no direct replay with true re-evaluation, no GP multitask or transfer competitor, and no neural or RL scheduler.

Ablations: four ablations cover the three components. The full method is not statistically significantly better than Equal-weight or Focused-weight on AP and only marginally better on other metrics, so the weighting component is weakly supported. The one-stage and raw-surrogate ablations are more convincing.

Surrogate analysis: the distance and MAE analyses support the threshold idea. The weight-estimation validation is a nice check, but only one order is shown for the surrogate figures.

Forgetting metrics: LLSGP is not the best on AF or AMF, where LLR-BC has lower mean values. The authors attribute this to LLR-BC's low plasticity, which is plausible but not tested. The abstract's claim of mitigating forgetting is therefore only partly borne out in statistical terms.

Open questions:

- The sequence is short (4 tasks).

- The two-stage switch threshold, window and τ are tuned in supplementary sensitivity analyses.

- The "first-tasks" results are all measured against a multi-task baseline that uses the same budget.

- The ~6% extra training time is reported, but surrogate construction is assumed to happen in idle time.

- Rule-size bloat is noted but not addressed.

Potential Impact

Practical: the immediate audience is the GP/hyper-heuristic/evolutionary scheduling community, where this would be a useful reference and baseline for lifelong GP. Real-world uptake in manufacturing is speculative because the experiments are simulated, with synthetic tasks and a single objective (mean weighted tardiness).

Broader: the ideas are transferable to other GP-evolved heuristics and to lifelong evolutionary computation. These include distance-gated surrogates, similarity-based replay weights, and stagnation-triggered stage switching. They are modest, domain-agnostic tricks. They are not likely to reach the mainstream continual-learning community, which is dominated by neural methods.

Timeliness & Relevance

Continual learning and non-stationary optimisation are hot topics, and the "lifelong" framing is timely for evolutionary computation. The niche (GP for DFJSS) is mature but incremental, and the paper fits a clear gap. Interest in interpretable rules is a mild plus.

Strengths

Clear problem framing with sensible metrics adopted from the continual-learning literature.

A coherent, well-motivated pipeline with an ablation for each component.

Thorough statistics over six orders and 30 runs.

Diagnostic analyses of weights, switching points, surrogate distances and rule size.

Code link provided (anonymous), and settings are fully tabulated.

Limitations

Only four tasks and synthetic simulation; no real-world or larger scale tests.

Lower forgetting than LLR-BC is not demonstrated; the gains are mainly in AP, i.e. a better stability–plasticity trade-off.

Hand-set thresholds (τ=5 in a 40-dimensional integer space, ε=0.1%) may not transfer across problems.

Weak significance for the task weighting component.

Surrogate memory grows with the number of tasks (P samples per task) and the combined fitness requires evaluating n×P candidates against all previous surrogates, so scalability to long sequences is unaddressed.

The multi-task baseline is used both as the upper bound and as normaliser, yet LLSGP's normalised values are still ~30% off it.

Incremental relative to the authors' own prior surrogate-assisted GP work.

Overall

A solid, competently executed domain-specific contribution that opens a modest new direction (lifelong GP for scheduling). It is likely to be cited mainly within the evolutionary scheduling community.

Rating:5.5/ 10

Significance 4.5Rigor 6Novelty 5.5Clarity 6.5

Generated Nov 2, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 16/16

5.5/10 impact

Competent, well-evaluated first treatment of forgetting in GP for DFJSS, but incremental over prior surrogate-assisted GP work, small task sequence, and niche audience.