{"id":"6664a1d7-291a-4cb7-ab6c-d91b161f1b4b","arxiv_id":"2509.11016","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A DDQN policy trained offline on CEC2010 constrained problems selects per-generation DE mutation strategies and constraint-handling techniques, and the paper reports it outperforms six classic DE variants on unseen benchmarks.","lead":"Researchers trained an AI agent that, at every generation of a Differential Evolution optimizer, picks the mutation strategy and the constraint-handling rule, after training on a set of constrained benchmark problems. The agent is then tested, without further training, on other benchmark suites and reported to beat six established DE variants on many problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No computational budget or resource-equivalence is reported; without equal max-FE/population settings, the claimed zero-shot superiority over six DE baselines is unfalsifiable.","rationale":"The reader's rejection is justified, and the stated weakest assumption (limited state/action design) is a real limitation of the claimed foundation-model scope. However, I find the more immediately load-bearing gap to be the missing computational-budget reporting and resource-equivalence control. The paper's Section V admits the state/action flexibility issue, but the empirical comparison itself is the foundation of the zero-shot claim. Without knowing whether SuperDE and the six baselines used the same number of function evaluations, the same population size, and the same termination rule, the reported performance gaps in Tables II, IV, VI, and X cannot be attributed to the learned policy. This is an omitted methodological support, not a disagreement with current consensus, and it is directly testable by asking for code/config and rerunning with fixed budgets. I therefore maintain the reader's REJECT verdict: the central claim is presently unsupported, though a revised version with code, full hyperparameters, and budget-fair benchmarks could become conditionally acceptable.","tokens_in":31915,"tokens_out":6373,"duration_ms":84277,"concrete_test":"Obtain from the authors (or from the released repository) the exact termination condition, max FE budget, population size, F, CR, epsilon schedule, and baseline configurations, then rerun all RQ1/RQ2 experiments with one common FE budget and otherwise identical environment settings. If the reported +/=/− counts and mean objective/CV values persist under equal budgets, the resource-equivalence concern is settled; if they shrink or reverse, the central zero-shot superiority claim does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A defines the evaluation metric but never states the computational budget (max function evaluations or generations), population size, F, CR, or termination rule for SuperDE or for any baseline. Algorithm 1 leaves termination abstract as \"done←algorithmTermination(Ij)\" (line 21). The RQ1/RQ2 tables compare average objective/CV values, but if SuperDE is allowed more function evaluations or a different stopping criterion, the apparent superiority in Tables II, IV, VI, and X could be an artifact of unequal resources rather than evidence for the learned component-selection policy. The central claim is that the pretrained DDQN generalizes zero-shot with superior optimization performance; that claim requires a fair, resource-equivalent comparison against the DE variants. The authors' own limitation about the manual state/action design is a genuine generalization concern, but it is secondary: even on the tested benchmarks, the comparison is not yet interpretable without reporting or controlling the budget. The absence of code and full hyperparameters compounds this, making the headline empirical result impossible to verify.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SuperDE, a Double Deep Q-Network (DDQN) controller that, at each generation of a Differential Evolution (DE) run, selects one of four mutation strategies and one of seven constraint-handling techniques. The controller is trained offline in a meta-learning fashion on CEC2010 constrained problems, and is then applied zero-shot to other benchmark suites (CEC2010 with different optima, CEC2017, G2000, BBOB2022). The authors claim that this learned per-generation component configuration significantly outperforms six DE-based baselines and also outperforms ablated variants (fixed MS, fixed CHT, random configuration). The paper includes convergence plots, Wilcoxon signed-rank test summaries, and an appendix table for BBOB2022.","tokens_in":32250,"tokens_out":7723,"duration_ms":92453,"significance":"If the empirical claims were fully supported, the paper would make a useful contribution to automated algorithm configuration: it demonstrates that a relatively small DDQN can be trained across a set of COPs and then used without retraining to select DE components on unseen benchmarks. The strengths are the clear MDP formulation, the inclusion of several independent held-out suites (CEC2017, G2000, BBOB2022), and an ablation study that attempts to isolate the contribution of each configured component. The authors also honestly list limitations (manual state/action design, limited training corpus, no discrete/combinatorial problems). However, the current experimental reporting is not yet sufficient to support the headline claim of zero-shot superiority: the computational budget is not specified, and the statistical significance tests are computed on a favorable subset of problems. These are load-bearing issues, not presentation details.","major_comments":[{"comment":"The evaluation setup never states the computational budget. The termination condition is left abstract as 'done←algorithmTermination(Ij)' (Algorithm 1, line 21), and no maximum number of function evaluations, population size, F, CR, or stopping rule is reported for SuperDE or for any competitor. Without resource-equivalent comparisons, the lower objective values and CVs in Tables II, IV, VI, and X could be an artifact of unequal budgets rather than evidence for the learned policy. This is central to the paper's claim of zero-shot superiority, and the setup must be specified and controlled.","section":"Section IV-A and Algorithm 1 (line 21)"},{"comment":"The '+/-/=' counts appear to exclude comparisons in which one algorithm fails to find a feasible solution. For example, in Table II, C2010E F9: SuperDE returns NaN while CMODE, IepsilonJADE, and C2ODE find feasible solutions. This should be counted as a '+' for those competitors, but the summary row does not appear to include it. The same issue occurs in Table IV (e.g., rows where SuperDE is NaN but a competitor has a value). Excluding such cases biases the Wilcoxon test in favor of SuperDE. The authors should either include these cases as losses, or explicitly state and justify that the test is restricted to mutually feasible problems, and qualify the 'significantly outperforms' claim accordingly.","section":"Tables II and IV (Wilcoxon summaries)"},{"comment":"The paper is not reproducible from the information provided. The training procedure lacks learning rate, optimizer, epsilon-greedy schedule, target network update frequency (Eq. (11) is not instantiated with a value), number of epochs Epo, number of instances M, and meta-learning loop details. The DE parameters F and CR for SuperDE and for all baselines are never given. Some state features are also under-specified: feature 14 relies on DBSCAN but the DBSCAN parameters are not reported, and feature 15 ('ratio of feasible boundary crossings') is not defined. No source code is provided. Given that the contribution is exclusively empirical, these omissions are major.","section":"Sections III-C, III-D, III-E, and IV"}],"minor_comments":[{"comment":"The definition of 'C2010E' is vague: 'with different optima' is not explained. Please clarify how the optima are shifted and whether this creates genuinely different instances from the CEC2010 training set.","section":"Section IV-A"},{"comment":"Features 14 and 15 lack precise definitions. For DBSCAN-based clustering, the radius and minimum samples parameters should be stated; the computation of 'feasible boundary crossings' should be formalized.","section":"Section III-C.1"},{"comment":"The candidate pool is small (4 MSs and 7 CHTs) and fixed. The authors acknowledge this limitation in Section V, but the abstract's 'foundation model' terminology seems disproportionate for a policy with a 28-action discrete space trained on one benchmark suite. A more measured term such as 'pretrained controller' would be more accurate.","section":"Section III-C.2"},{"comment":"Some table entries have no +/-/= marker when the value is NaN (e.g., Table II), which makes the summary rows difficult to verify. Figure 4 is hard to read in the text; the legend and line colors should be improved. References [24] and [34] appear to be duplicates of the same Kiran et al. paper and should be merged.","section":"Tables and figures"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the paper contains a substantial experimental study, but the missing computational budget and the selective Wilcoxon test are serious. I recommend major revision rather than reject because these issues are fixable: the authors can add a precise budget definition, re-run or re-report the comparisons under equal resources, and correct the statistical analysis. If, after revision, the resource-equivalence cannot be demonstrated, the paper should not be accepted. I would also ask the editor to consider whether the 'foundation model' framing is appropriate for the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The core idea—train a DDQN offline on a corpus of COPs to pick mutation strategy and constraint handling each generation, then apply it zero-shot to unseen problems—is a genuine extension of the per-problem DRL adaptation literature. The 22-feature state and 28-action pool are well-motivated, and the paper is honest in Section V that the manual state/action design and the small training corpus limit generality. The authors also do a decent job of placing their work relative to Ming, Wang, Tian, and Luo.\n\nThe empirical section is where the paper fails to convince. The stress-test is on target: there is no computational budget anywhere. Algorithm 1 leaves termination abstract, and the tables report average minimum objectives without stating max function evaluations, population size, or even F/CR for SuperDE or any baseline. Without resource-equivalence, the claimed superiority is unfalsifiable—the DDQN might just be running longer.\n\nThe second problem is the statistics. The Wilcoxon summaries are computed on a subset of problems, excluding cases where SuperDE returns NaN but a competitor finds a feasible solution. C2010E F9 in Table II is a clean example: SuperDE fails, CMODE/IepsilonJADE/C2ODE succeed, and that comparison is silently dropped from the +/-/= counts. That biases the win rates.\n\nThird, no code is released, and many hyperparameters are missing. For an RL-based configurator, that's a reproducibility gap.\n\nTo give credit: the raw tables show some real strengths. SuperDE finds feasible solutions on C2017 F6 and F10 where all six baselines fail, and its CV values on infeasible problems are often competitive. So the method is not obviously broken. But the paper as presented overclaims.\n\nI'd send this to peer review, with the expectation of major revision. The reviewers should demand a full experimental setup (budget, termination, hyperparameters) and a statistical treatment that counts NaN as a loss for SuperDE. If the authors can deliver that, the result might be publishable. As is, I wouldn't cite the headline claim.\n\nFor reading group, maybe—the budget issue is a good discussion point, but the table-heavy results will slow things down.","headline":"SuperDE has a worthwhile idea but the empirical comparison is not interpretable as written—no budget reported, and the statistics skip SuperDE's failures.","tokens_in":32719,"tokens_out":3832,"would_cite":false,"duration_ms":42359,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep reinforcement learning policy, trained on a diverse set of constrained problems, can automatically select per-generation mutation and constraint-handling components for Differential Evolution and generalize zero-shot to unseen proble","keywords":["deep reinforcement learning","differential evolution","constrained optimization","component auto-configuration","zero-shot generalization","meta-learning","DDQN","constraint handling technique"],"falsifier":"A decisive test: generate a family of constrained optimization problems with a tunable parameter that changes the geometry of the feasible region (e.g., disconnectivity or the ratio of active constraints) and hold out a subset. If SuperDE's average objective/CV on the held-out instances is not better than the best fixed component pair chosen in hindsight, or if a random-configuration baseline performs equally well on problems where the oracle-optimal action lies outside the pool, then the claim that the learned policy generalizes zero-shot fails.","tokens_in":31821,"feed_emoji":"🧬","tokens_out":7930,"duration_ms":85906,"temperature":0.7,"pith_summary":"The paper introduces SuperDE, a deep reinforcement learning framework that automates the configuration of Differential Evolution (DE) components for constrained optimization problems (COPs). Its central claim: after offline meta-training on a diverse set of benchmark COPs, a Double Deep Q-Network policy can, at each generation, select one of four mutation strategies and one of seven constraint-handling techniques from a 22-feature population state, and this policy transfers to unseen problems in a zero-shot manner. On CEC2010, CEC2017, G2000, and BBOB2022 benchmark suites, SuperDE is reported to outperform six DE-based algorithms in finding feasible solutions and minimizing objective values or constraint violations. The paper also argues through ablations that simultaneously configuring both components yields a '1+1>2' effect, and that the offline-trained policy avoids the computational cost of per-problem online tuning.","feed_headline":"SuperDE: zero-shot RL beats hand-tuned DE on constrained problems","feed_subtitle":"A DDQN agent, trained offline, configures DE every generation and beats tuned variants on unseen benchmarks.","key_machinery":"The core mechanism is an MDP that turns algorithm configuration into a sequential decision problem: state = 22 population features; action = 2D discrete tuple (mutation strategy index, CHT index) from a pool of 4 MSs (DE/rand/1, DE/best/1, DE/cur-to-best/1, DE/rand-to-best/1) and 7 CHTs (death penalty, weighted penalty, two feasibility rules, tournament, epsilon feasibility, stochastic ranking); reward = phase-dependent scalar that prioritizes violation reduction when no feasible solution exists and balances objective improvement with feasible-ratio growth otherwise. A Double Deep Q-Network with target network, experience replay, and prioritized sampling learns Q-values from trajectories col","core_discovery":"On the paper's own terms, the central discovery is that per-generation component configuration for COPs can be learned as a Markov decision process and solved with deep reinforcement learning, producing a policy that generalizes across problem instances. The policy's input is a 22-dimensional vector of population statistics—feasibility ratio, fitness and objective extrema, constraint violation statistics, feasible-region cluster structure, correlation between objective and violation, and so on—and its output is one of 28 discrete actions, each selecting a mutation strategy and a CHT. Trained with a Double Deep Q-Network across many problem instances, this policy is applied without retraining","pith_inferences":["One untested corollary is that the learned policy might serve as a warm start for online DRL-based configuration on a new problem class, potentially reducing the number of evaluations needed to match problem-specialized algorithms.","The fixed action pool is the chief limiting factor; if the state representation were combined with generative or continuous actions (e.g., F and CR values, or epsilon schedules), the same architecture could cover a much larger configuration space.","The 22 features include structural descriptors like feasible clusters and boundary-crossing ratio; an ablation of these features could reveal which population signals drive the policy's decisions, making the model interpretable and possibly improving state design.","A natural stress test would be to apply SuperDE to problems with time-varying constraints or to a combinatorial constrained problem (which the paper lists as a limitation); the zero-shot claim only covers continuous COP benchmarks."],"forward_implications":["If SuperDE generalizes as claimed, a single pretrained policy can serve as an off-the-shelf configurator for a broad class of constrained optimization problems, replacing per-problem hand tuning.","The reported advantage over SuperDE1/2/3 implies that the coordinated choice of mutation and CHT is more than the sum of independent selections; practitioners should treat these components as coupled.","Because training is offline, the online overhead of SuperDE is just one small network forward pass per generation, which is negligible relative to the cost of function evaluations in expensive engineering problems.","The paper's ablation results suggest that even random configuration (SuperDE3) is competitive on some problems, indicating that pool diversity alone contributes part of the gain; the learned policy's value is in consistency across problems.","The paper states that the model can be fine-tuned online; if that is effective, users could adapt the pretrained policy to a specific domain with only a small number of additional gradient updates."],"fun_headline_variants":["SuperDE: RL configures DE per generation, zero-shot on new problems","Deep RL auto-tunes DE components for constrained optimization","SuperDE beats hand-tuned DE using DRL-guided component selection","Zero-shot RL configures DE each generation, generalizes across COPs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the 22 hand-crafted population features plus the fixed pool of 4 mutation strategies and 7 constraint-handling techniques suffice to represent the optimal configuration decision for any unseen constrained problem; if an unseen problem's best per-generation action falls outside this discretization, the zero-shot guarantee collapses (a limitation the paper acknowledges in its conclusion).","fun_headline_variants_meta":{"raw":{"variants":["SuperDE: RL configures DE per generation, zero-shot on new problems","Deep RL auto-tunes DE components for constrained optimization","SuperDE beats hand-tuned DE using DRL-guided component selection","Zero-shot RL configures DE each generation, generalizes across COPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000373,"raw_usage":{"total_tokens":1821,"prompt_tokens":725,"completion_tokens":1096,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1030}},"tokens_in":469,"tokens_out":1096,"duration_ms":10231,"temperature":1.0,"reasoning_tokens":1030,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:13:53.644110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: generate a family of constrained optimization problems with a tunable parameter that changes the geometry of the feasible region (e.g., disconnectivity or the ratio of active constraints) and hold out a subset. If SuperDE's average objective/CV on the held-out instances is not better than the best fixed component pair chosen in hindsight, or if a random-configuration baseline performs equally well on problems where the oracle-optimal action lies outside the pool, then the claim that the learned policy generalizes zero-shot fails.","supporting_citations":[],"review_version":1}