{"id":"248d66ed-74c9-4564-9451-04698b79c19d","arxiv_id":"2505.16471","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper shows that a GNN-based RL controller that reads a graph of the current Pareto front can outperform static tuning and prior RL tuners on multi-objective scheduling problems.","lead":"This paper trains a reinforcement learning agent with a graph neural network to adjust evolutionary algorithm parameters, such as crossover and mutation rates, on the fly while solving multi-objective scheduling and routing problems. It reports improvements in solution quality, especially on five-objective scheduling problems, and shows the learned policies transfer to larger and more constrained instances.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Static AC baselines are tuned on the test instances, so the Penta-FJSP 'significantly better' results rest on a test-set-tuned comparison; a clean train/tune/test re-evaluation is needed.","rationale":"The reader's verdict is CONDITIONAL, and the rationale already notes the protocol error that static baselines are tuned on the test instances. My stress-test identifies this as the single most load-bearing concern about the central empirical claim, because the abstract's and RQ1's 'significantly better' statements are directly supported by Table 1, whose validity is compromised by test-set leakage into baseline tuning. The reader's weakest_assumption, however, concerns the graph-state representation; that is a different concern about the mechanism behind the method, not about whether the headline outperformance exists. I therefore partially agree with the reader: the protocol error is present in the reader's rationale, but it was not chosen as the weakest assumption. My concrete test—re-tuning baselines on a disjoint set and recomputing the Penta-FJSP comparison—would settle whether the reported advantage is genuine or an artifact. The verdict stays CONDITIONAL because either way the manuscript needs revision: first to fix the protocol, and second to add the graph-ablation evidence that the reader's weakest_assumption calls for.","tokens_in":19472,"tokens_out":8622,"duration_ms":73909,"concrete_test":"Re-run irace and SMAC3 with the same tuning budget but using a tuning set disjoint from the test set (e.g., the 100 training instances per size, or a newly generated validation set), then evaluate on the held-out 100 test instances for Penta-FJSP 25j5m. Recompute Table 1 mean/max hypervolumes and the Wilcoxon rank-sum p-values vs GS-MODAC. If GS-MODAC remains significantly better than the best properly-tuned baseline, the central claim survives; otherwise the reported advantage is an artifact of test-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 (Baselines) states that SMAC3 and irace are tuned 'using the generated test instances for each given instance size'; Table 1 then evaluates all methods on those same test instances. This gives the two static baselines access to test-set information during configuration, allowing them to overfit to the evaluation instances. The headline quantitative claim—GS-MODAC being 8.2% (mean) and 5.7% (max) better than the best baselines on Penta-FJSP 25j5m—is therefore not a comparison under a standard, deployable protocol. A correct evaluation would tune static baselines on a training/validation split and test on held-out instances. Without this, the abstract's 'outperforms' claim cannot be reliably assessed, and the reported margins may change (in either direction) once the baseline configurations are no longer specialized to the test set. This is the most load-bearing weakness because the central empirical claim rests on Table 1, and RQ1's 'significantly better' statement depends on the integrity of this comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GS-MODAC, a deep reinforcement learning method for dynamic algorithm configuration of multi-objective evolutionary algorithms. The method represents the current population's solutions in objective space as a graph whose nodes carry normalized objective values and whose edges connect solutions within the same Pareto front; a GCN encodes this graph, and PPO trains a policy that outputs continuous crossover and mutation parameters (or MOPSO coefficients) at each generation. The reward is based on squared hypervolume improvements relative to an initial nadir point and an approximate ideal point. Experiments cover bi-, tri-, and penta-objective FJSP variants and bi-objective CVRP, with comparisons against NSGA-II with default parameters, SMAC3, irace, and MADAC, plus transfer tests to larger instances, more constrained scheduling variants, different objective sets, and a different MOEA.","tokens_in":19684,"tokens_out":6231,"duration_ms":53359,"significance":"If the reported results hold, this is a useful contribution to dynamic algorithm configuration for multi-objective combinatorial optimization. The paper introduces a graph-based state representation that is independent of the number of objectives, an instance-agnostic reward, and demonstrates transfer across problem sizes, objective counts, constrained variants, and to a second MOEA. The code is publicly available, which is a reproducibility strength. The main empirical claims, however, are currently weakened by a baseline tuning protocol that gives SMAC3 and irace access to the test instances, and by the absence of an ablation isolating the contribution of the graph structure itself.","major_comments":[{"comment":"The static baselines SMAC3 and irace are tuned 'using the generated test instances for each given instance size' and then evaluated on those same instances. This is not a standard deployable evaluation protocol: the two static tuners receive test-set information during configuration, so their reported performance is optimistically biased. Since RQ1 and the abstract's 'outperforms' claim rest on Table 1, the headline margins (e.g., Penta-FJSP 25j5m: mean +8.2%, max +5.7% over the best baselines) cannot be taken at face value. The authors should retune SMAC3 and irace on training/validation instances and evaluate on held-out instances, and should report whether the differences persist under that protocol.","section":"Section 4 (Baselines); Table 1"},{"comment":"The graph-based state is the central novelty, but the ablations do not isolate the contribution of the graph structure. Removing the budget feature and reducing the GCN to one layer produce performance changes of roughly 0.8-3.4%, and replacing GCN with Transformer or GAT yields similar results, yet no ablation compares the proposed graph state to an edge-free node embedding, a fully connected graph, or a hand-crafted vector of objective-space statistics. Consequently, the claim that the graph topology rather than the GNN embedding of normalized objective values drives the reported gains is not supported. An ablation that varies the graph construction while keeping the encoder and reward fixed would address this directly.","section":"Section 3.1 (States); Appendix E"},{"comment":"The paper marks results as 'significantly outperform[ing]' via a Wilcoxon rank-sum test (p < 0.05) but does not state the unit of the test (per instance, per run, or per instance-mean), whether the test is paired, or whether any multiple-comparison correction is applied across the many entries in Table 1. Several reported improvements are within one standard deviation (e.g., Bi-FJSP 10j5m: 3.92 x 10^4 vs 3.89 x 10^4 with standard deviations around 2 x 10^3). The authors should report the test unit, effect sizes, and correction procedure, or restrict significance claims to comparisons where such a test is clearly defined.","section":"Section 4.1; Table 1"}],"minor_comments":[{"comment":"The text refers to 'CVPR' in one place; this should read 'CVRP'.","section":"Section 4 (Baselines)"},{"comment":"Number formatting is inconsistent (e.g., '1.000,000 steps' and '10.000 runs'); please use uniform separators for readability.","section":"Section 4 (Training)"},{"comment":"The figure caption does not state which instance is shown, whether the curves are averaged over runs, or whether error bars or confidence bands are included; please add this information.","section":"Figure 2"},{"comment":"The text says GS-MODAC finds a significantly higher number of 'best' solutions, but no significance test is reported for these additional metrics, and the table mixes mean/min/max rows without explaining which quantity is used for each claim.","section":"Appendix C, Table 5"},{"comment":"The comparison with P-MOCO uses only the best HV value, not mean and standard deviation, and is limited to CVRP; the conclusion about P-MOCO's inferior generalization should be presented as preliminary given this restricted evidence.","section":"Appendix G"}],"recommendation":"major_revision","confidential_remarks":"The test-set tuning of SMAC3 and irace is the most serious issue and is likely fixable by re-running the tuners on training/validation instances and re-evaluating on held-out instances. I recommend that the editor require this re-evaluation before acceptance, together with a clearer statistical reporting protocol. The paper fits the journal's scope and the code availability is a strength. If the re-evaluation confirms the reported margins, I would support publication after the remaining revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper's genuine contribution is the graph state representation for dynamic MOEA configuration—nodes are normalized objective values, edges link solutions within the same non-dominated front, and a GCN produces the state embedding. That is a real improvement over MADAC's hand-crafted scalar features, and the reward based on a double-budget HV_ideal is nicely instance-agnostic. Both are worth taking seriously.\n\nWhat it does well: the method is clearly specified, the ablation study (budget feature, one GCN layer, transformer/GAT alternatives) is useful, and the MOPSO deployment plus the cross-size and cross-variant generalization experiments add value. Code is public. The writing is straightforward.\n\nThe soft spot is the evaluation protocol. In Section 4, SMAC3 and irace are tuned on the generated test instances and then evaluated on those same instances. That's test-set tuning, which invalidates the clean comparison and makes the headline 'significantly better' claim for Penta-FJSP hard to trust. Note the bias direction: the static baselines are given the test instances, so they get an unfair advantage; GS-MODAC still wins on Penta, which is encouraging, but the reported margins and p-values are not based on a standard train/tune/test split. This needs a re-run. Also, many of the improvements are within one standard deviation; only the Penta-FJSP results look decisive. The abstract's blanket 'outperforms' is stronger than the data support, especially since the matched-size P-MOCO comparison in Appendix G actually shows P-MOCO beating GS-MODAC on the size it was trained on—the generalization story is fine, but the claim should be qualified.\n\nMinor: the graph representation edges only within the same front; no ablation isolates whether the graph structure itself matters, rather than just the GNN embedding of normalized values. Not fatal, but it weakens the 'graph-support' story.\n\nWho is this for? Anyone working on DRL-based algorithm configuration or on learning to tune MOEAs for scheduling/routing. It deserves a proper referee even with the protocol issue—the idea is new enough and the experimental breadth is enough that a careful re-evaluation with a clean split is exactly what peer review should demand. I'd advise asking for a revised version with a clean split, full significance stats, and tempered claims.","headline":"A real graph-state contribution for DRL-based MOEA configuration, but the headline win over static baselines is weakened by test-set tuning.","tokens_in":20225,"tokens_out":3647,"would_cite":true,"duration_ms":30886,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C27","90C29","68T20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep reinforcement learning agent that reads the current solution population as a graph can dynamically configure multi-objective evolutionary algorithms, beating static tuning and earlier DRL-based configuration and generalizing across…","keywords":["dynamic algorithm configuration","multi-objective combinatorial optimization","graph neural network","deep reinforcement learning","multi-objective evolutionary algorithm","hypervolume reward","flexible job shop scheduling","capacitated vehicle routing"],"falsifier":"Retrain the same pipeline on Penta-FJSP with 25 jobs and 5 machines using the same normalized node features but with the graph edges removed, so the state is just the pooled node features plus the budget scalar; if the mean test hypervolume stays near $5.62\\times 10^{12}$, the graph structure is not doing the work, whereas a drop toward the SMAC3 or irace level would confirm the graph state as the operative component.","tokens_in":19269,"feed_emoji":"🧬","tokens_out":10799,"duration_ms":80909,"temperature":0.7,"pith_summary":"GS-MODAC claims that the state of a multi-objective evolutionary search is best represented as a graph of the current solutions in objective space, and that a deep reinforcement learning agent trained on that graph can choose the algorithm's parameters from one generation to the next more effectively than static tuning or earlier DRL-based configuration. The target problems are combinatorial—flexible job-shop scheduling and capacitated vehicle routing—where earlier dynamic configuration approaches, built for continuous benchmarks, lose traction. The paper's proposed answer is to let a graph neural network learn the state from normalized objective values and Pareto-front structure, with an instance-agnostic reward based on hypervolume improvement, so no hand-crafted state features are needed. If the results hold, automatic dynamic parameter control becomes practical for a class of NP-hard problems where tuning is currently expensive and static settings waste performance.","feed_headline":"Graph-based tuner beats static configuration on multi-objective search","feed_subtitle":"A graph-neural-network state and hypervolume reward let one trained controller lift results and transfer across sizes.","key_machinery":"The central object is the graph state: each solution is a node whose features are its normalized objective values, and edges connect only solutions lying on the same non-dominated front, so the graph carries the hierarchy of Pareto fronts without any hand-crafted convergence or diversity metrics. Two graph convolutional network layers followed by global mean pooling turn the population into one embedding; this embedding is concatenated with the normalized remaining generation budget and passed through a linear layer to produce the means of the continuous action distributions for PPO. The reward is built on the hypervolume indicator: when the current hypervolume beats the best seen so far, the agent receives the difference between the squared percentage improvements over the initial hypervolume, with the ideal hypervolume obtained from a doubled-budget run, so reward values are comparable across instances of different scale.","core_discovery":"On its own terms, GS-MODAC discovers that the evolving population of a multi-objective heuristic can be encoded as a graph—one node per solution holding normalized objective values, edges linking solutions on the same Pareto front (the set of mutually non-dominated solutions)—and that a graph convolutional network embedding of this graph, together with a remaining-budget scalar, is a sufficient state for a PPO agent to set crossover and mutation rates (or particle-swarm coefficients) that beat static configurators and a multi-agent DRL baseline. The reward pays off only when the population's hypervolume improves, and it equals the difference between squared percentage improvements relative to the initial and ideal hypervolumes, which makes the reward scale-free across instances and emphasizes the harder late-stage gains. On the five-objective, 25-job flexible job-shop problem, the trained policy's mean and maximum hypervolumes are 8.2% and 5.7% better than the strongest baselines, and the same policy transfers to larger vehicle-routing instances, an assembly-and-setup-time variant, and objective pairs not seen in training.","pith_inferences":["A direct ablation the paper does not run would isolate whether the graph edges matter: retrain with the same normalized node features but no edges, or with edges between all pairs of solutions, and compare test hypervolumes on Penta-FJSP.","The graph-state idea is not tied to evolutionary computation; any iterative multi-objective method that maintains a set of candidate solutions—local search, simulated annealing, or beam search—could use the same normalized-objective graph as the state for a controller.","The reward's dependence on a doubled-budget run to obtain the ideal point suggests a practical extension: replace that fixed reference with an adaptive or learned estimate so the method applies when extra compute is unavailable.","If the size-transfer results generalize, a single policy trained on a mixture of small instances could serve as a default parameter controller for an entire problem family, removing the need to retune per instance size."],"forward_implications":["Practitioners can drop hand-designed state features, because the graph state is constructed directly from normalized objective values and Pareto-front membership.","A trained policy transfers to larger instances of the same problem with only a slight performance decline, and still matches or beats the best baselines.","A policy trained on one scheduling setting also transfers to a more constrained variant with assembly constraints and sequence-dependent setup times, outperforming static configurators tuned on that variant.","The framework is not tied to one base algorithm: it configures both NSGA-II and MOPSO.","The added cost is small—state construction and policy inference account for about 2.0% of total runtime on the smallest scheduling instances—so the dynamic control is cheap relative to the search itself."],"supporting_citations":[{"why":"Defines dynamic algorithm configuration as a sequential decision problem, the formal foundation GS-MODAC builds on.","marker":"(Biedenkapp et al., 2020)"},{"why":"The MADAC method is the closest DRL-based dynamic configuration baseline and the main comparison target for multi-objective performance.","marker":"(Xue et al., 2022)"},{"why":"Provides the irace static configurator baseline that GS-MODAC must match or beat.","marker":"(López-Ibáñez et al., 2016)"},{"why":"Provides the SMAC3 Bayesian-optimization configurator baseline.","marker":"(Lindauer et al., 2022)"},{"why":"Defines the hypervolume indicator used both in the reward function and in evaluation.","marker":"(Zitzler & Thiele, 1998)"},{"why":"Supplies NSGA-II, the main multi-objective evolutionary algorithm whose parameters are configured.","marker":"(Deb et al., 2002)"},{"why":"Supplies PPO, the policy-gradient algorithm used to train the controller.","marker":"(Schulman et al., 2017)"},{"why":"Supplies the graph convolutional layers that embed the population graph into the state.","marker":"(Kipf & Welling, 2016)"},{"why":"Supplies the rule-of-thumb parameter ranges used for the action space and the vanilla NSGA-II baseline.","marker":"(Coello et al., 2007)"}],"fun_headline_variants":["GNN-tuned MOEAs beat static configs on multi-objective benchmarks","Reinforcement learning tunes evolutionary algorithms via graph state","Graph-based DRL configures multi-objective search to beat static","Trained GNN policy lifts MOEA results and transfers across sizes","Graph-to-action: DRL configures MOEAs with transferable gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a graph whose nodes are normalized objective values and whose edges only connect solutions that do not dominate each other carries enough information about the search state for the agent to choose good parameter updates; if the missing cross-front and absolute-scale information matters, the claimed generalization collapses.","fun_headline_variants_meta":{"raw":{"variants":["GNN-tuned MOEAs beat static configs on multi-objective benchmarks","Reinforcement learning tunes evolutionary algorithms via graph state","Graph-based DRL configures multi-objective search to beat static","Trained GNN policy lifts MOEA results and transfers across sizes","Graph-to-action: DRL configures MOEAs with transferable gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000598,"raw_usage":{"total_tokens":2778,"prompt_tokens":906,"completion_tokens":1872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":1779}},"tokens_in":522,"tokens_out":1872,"duration_ms":10704,"temperature":1.0,"reasoning_tokens":1779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:00:08.357484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same pipeline on Penta-FJSP with 25 jobs and 5 machines using the same normalized node features but with the graph edges removed, so the state is just the pooled node features plus the budget scalar; if the mean test hypervolume stays near $5.62\\times 10^{12}$, the graph structure is not doing the work, whereas a drop toward the SMAC3 or irace level would confirm the graph state as the operative component.","supporting_citations":[{"cited_title":"F., Eimer, T., Hutter, F., and Lindauer, M","cited_arxiv_id":null,"evidence_quote":"Defines dynamic algorithm configuration as a sequential decision problem, the formal foundation GS-MODAC builds on."}],"review_version":1}