{"id":"4cfb3008-46f4-41e1-a694-1671140fc14a","arxiv_id":"2505.14610","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The MMD-Newton method adds a fast, derivative-based refinement stage to evolutionary multi-objective optimization, improving Pareto front accuracy on most tested benchmarks.","lead":"A new refinement method for multi-objective optimization minimizes the MMD distance between a candidate solution set and a reference set, using a Newton step that needs derivatives of the objectives. Combined with evolutionary algorithms, the hybrid improved accuracy on 28 of 33 benchmark test cases compared with the evolutionary algorithm alone.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hybrid's central robustness claim — that MMD's diversity term rescues an imperfect reference set — is unsupported because R is always built from the very population Y0 being refined; when Y0 is subregion-biased, R inherits the bias and MMDN cannot recover, exactly the paper's one loss (MOEA/D…","rationale":"I read the paper in good faith: the algorithm is well-specified, the chain-rule derivatives (Eqs. 12-17) are correct, the Hessian preconditioning (Section C.1) is standard, and the budget-equivalence accounting (Section C.3) is a reasonable attempt at fairness. The empirical table is the paper's strongest asset: 28/33 significant wins with Holm-Sidak correction is a real result even if the absolute improvements are modest.\n\nThe central claim, however, is two-part: the empirical claim (hybrid beats EA alone) is supported, while the mechanism claim (MMD's diversity term rescues imperfect reference sets, the paper's stated motivation) is the load-bearing part least secure. The reference set is constructed from Y0 itself (Algorithm 1 lines 5-9, Section B), so the target encodes whatever bias the warm start has. The paper's own Section 7 admits this: 'the performance of MMDN can be affected by the quality (e.g., the diversity) of the Pareto approximation set of MOEA, as shown by the DTLZ7 problem with MOEA/D case in the previous section.' That admission, combined with the absence of any systematic recovery study (Fig. 1 is a single illustration with IGD of about 0.022), makes the robustness claim unverified rather than merely conservative.\n\nI also examined the theory for internal consistency. Lemma 1's 'iff' is not justified: from Eq. (24), the derivative is an integral against the spectral measure G, and the integral vanishing does not imply the integrand vanishes pointwise (the 'only if' direction), since the spectral integral can vanish by symmetry even when the bracketed sine term is nonzero; only the 'if' direction is immediate. This is a real but secondary flaw, since Theorem 1 and Theorem 2 do not appear to rely on the converse.\n\nMinor issues: the per-problem kernel and length-scale grid search (Section C.2, Table 2) selects hyperparameters on the benchmark problems themselves, so the wins are not fully out-of-sample; the code link (anonymous.4open.science/r/HypervolumeDerivatives-5287) appears to point to a different project, hurting reproducibility; and the Algorithm 1 stopping condition (line 11) has an inverted inequality, iterating while the gradient is small and stopping when it is large.\n\nOn balance, the reader's CONDITIONAL verdict is appropriate. The proposed truncation experiment would settle whether the abstract's robustness claim survives; if it fails, the paper should be revised to claim only local refinement of well-distributed warm starts, with the diversity-recovery claim removed or heavily scoped.","tokens_in":21049,"tokens_out":16880,"duration_ms":113545,"concrete_test":"Systematic bias-injection experiment: fix a connected-front problem (DTLZ1) and a disconnected-front problem (ZDT3 or DTLZ7), the three baseline MOEAs, N1=300, N2=5. Run the MOEA, then deliberately truncate Y0 to a subregion (e.g., keep only the half with smallest f1 for bi-objective cases; remove one corner of the simplex for DTLZ1). Build R from the truncated Y0 exactly as in Section B, run MMDN, and over 30 runs measure the final objective span of Y and the Delta-2 metric. If the final span does not extend substantially beyond the truncated subregion (or Delta-2 does not beat the truncated warm start), the diversity term does not rescue a biased R, the DTLZ7 failure is the rule rather than the exception, and the abstract's robustness claim must be scoped down or dropped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim has two load-bearing parts: (a) minimizing MMD^2(Y,R) improves coverage through a diversity term even when R is imperfect (Section 1, Fig. 1), and (b) the hybrid MMDN+MOEA beats EA alone on the same budget (Table 1). Both rest on the reference set construction: Algorithm 1 lines 5-9 and Section B build R from Y0, the MOEA's own final population, by interpolation, k-means reduction, and a delta-shift into the utopian region. R is therefore a smoothed, shifted copy of the very set being refined.\n\nConsequence 1 (robustness): if Y0 covers only a subregion of the front, R inherits that span. The MMD cross-term -2<psi(Y),psi(R)> then pulls every point toward R's subregion; points outside it feel exponentially small attraction (Gaussian kernel), while the self-term repulsion keeps spreading them, off the missing part of the front rather than onto it. The paper's only loss, MOEA/D on DTLZ7 (Table 1), is precisely this failure, and Section 7 concedes it.\n\nConsequence 2 (attribution): in the 28 wins, Y0 from a 300-iteration MOEA plausibly already spans the front, so R encodes the correct support and MMDN's improvement may be ordinary local Newton refinement of a good warm start (quadratic convergence of the set-valued Newton step, Eq. 10), not the diversity mechanism. Fig. 1 is the sole direct evidence for recovery, a single unquantified illustration with final IGD of about 0.022 on a flat simplex front.\n\nThe load-bearing assumption is therefore: the warm-start Y0 is diverse enough that R inherits the true front's span. This is not proven, is contradicted by the DTLZ7/MOEA/D row, and is exactly the premise the paper's motivating claim (recovery from imperfect R) needs most.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using maximum mean discrepancy (MMD) between a finite Pareto approximation set Y=F[X] and a reference set R as an objective for numerical multi-objective optimization. It derives analytical gradient and Hessian expressions for MMD with respect to the decision variables, formulates a set-oriented Newton method (MMDN), and analyzes a first-order stationary condition and Hessian eigenvalue bounds. The authors then hybridize MMDN with MOEAs by running an MOEA for 300 iterations, constructing R from the MOEA's final population Y0 via interpolation, k-means clustering, and a utopian shift, and running five MMDN iterations. They benchmark the hybrid against the same MOEA alone under an equivalent function-evaluation budget on 11 problems with three MOEAs, reporting 28 wins, 4 ties, and 1 loss. The central claimed advantage is that MMD's self-term maintains diversity, so the method can improve coverage even when R is imperfect.","tokens_in":21499,"tokens_out":6625,"duration_ms":68082,"significance":"If the central claims hold, the paper would make a useful contribution: it provides explicit derivative formulas for a set-based MMD objective, a Newton-type refinement strategy, and a practical hybrid that appears to improve several MOEA baselines. The analytical gradient/Hessian derivations in Section 4.1 are standard but useful, and the reported code link supports reproducibility. The empirical comparison is extensive for a paper of this length, and the win/tie/loss pattern across 33 configurations is concrete evidence that the hybrid often helps. However, the main theoretical lemma on stationary conditions is stated as an equivalence that is not actually proven, and the core robustness claim about 'imperfect reference sets' is undermined by the fact that R is constructed from the very population being refined. These issues are load-bearing for the paper's stated contributions, so the manuscript needs substantial revision before the claims can be accepted as stated.","major_comments":[{"comment":"Lemma 1's 'if and only if' direction is not established. From the stationarity equation ∂/∂y_l MMD²(Y,R) = (2/µ²)∫_{R^k} ω B(ω) G(dω) = 0, with B(ω)=Σ_{α≠l} sin⟨ω,y_α-y_l⟩ − Σ_β sin⟨ω,r_β-y_l⟩, one cannot conclude that B(ω)=0 for every ω. A nondegenerate, or even strictly positive, spectral measure G does not prevent cancellations: an odd integrand can integrate to zero over a symmetric measure without vanishing pointwise. Only the 'if' direction (pointwise identity implies zero gradient) follows directly. The lemma, and Remark 1 built on it, should be weakened to a sufficient condition or augmented with an additional assumption that rules out such cancellations.","section":"Section 4.2, Eq. (24), and Appendix A.1"},{"comment":"The central claim that MMD 'can improve the Pareto approximation set even when the reference R is imperfect' is not supported by the experimental protocol, because R is generated from the MOEA's own final population Y0: Y0 is interpolated, reduced by k-means, and shifted into the utopian region. If Y0 is confined to a subregion of the Pareto front, R inherits that span; the cross-term −2⟨ψ(Y),ψ(R)⟩ attracts points toward R's support, and the self-term repels points but does not direct them toward the missing part of the front. The paper's only loss in Table 1, MOEA/D on DTLZ7, is precisely this failure mode, and Section 7 concedes that the reference set and starting points are of low quality. To substantiate the diversity-recovery claim, the authors should test MMDN with a reference set independent of the initial population, or artificially restrict Y0 to a subregion and show that MMDN recovers full coverage. Without such an ablation, the observed wins are consistent with ordinary local Newton refinement of a good warm start rather than with the claimed diversity mechanism.","section":"Algorithm 1 lines 5-9, Section B, Section 7, and Table 1"},{"comment":"The eigenspectrum analysis applies to the objective-space block Hessian ∂²/∂y_m∂y_l MMD²(Y,R), but the Newton step and the preconditioning step in Algorithm 1 operate on the decision-space Hessian ∇² MMD²(X,R) defined in Eq. (17). The decision-space Hessian contains additional terms involving DF(x_m) and D²F(x_l), so its definiteness is not governed directly by the bounds in Corollary 1. The statement that preconditioning is necessary because 'the MMD Hessian is indefinite' therefore concerns a different matrix than the one being preconditioned. The authors should either prove corresponding bounds for the decision-space Hessian, or state explicitly that Corollary 1 concerns the objective-space MMD landscape and explain how that justifies the preconditioning of Eq. (17).","section":"Section 4.3, Corollary 1, Eq. (17), and Algorithm 1 line 12"},{"comment":"The claim of comparison 'with the same computation budget' is only approximate. The conversion from five MMDN iterations to an equivalent number of function evaluations uses measured CPU-time ratios for Jacobian and Hessian calls, but the reference set generation (DBSCAN, Delaunay triangulation, k-means, shifting) and the Hessian preconditioning/line-search overhead are not included in the budget. If these costs are non-negligible, the comparison is not strictly fair. At minimum, the wall-clock overhead of these steps should be reported, and the budget-conversion procedure should account for them or justify that they are negligible.","section":"Section 6, 'Fair comparison of MMDN to MOEA' and Section C.3"}],"minor_comments":[{"comment":"The while-loop condition reads 'while ||∇MMDM²(X_k)||₂ ≤ ε ∧ k < N2', which would execute the loop when the gradient norm is already below the tolerance; this should presumably be '> ε' or '≥ ε' if the intent is to stop once the gradient is small.","section":"Algorithm 1, line 11"},{"comment":"The Hessian formula mixes row-vector and tensor notation: the term (∂/∂y_l MMD²) D²F(x_l) δ_m^l should specify which index of the (1,2)-tensor is contracted, and the dimensions of the resulting block should be stated explicitly.","section":"Eq. (17)"},{"comment":"The averaged Hausdorff metric ∆_p used for performance assessment is not defined in the main text; since the paper cites [29] in the introduction, adding the definition or a one-line formula would make the results self-contained.","section":"Section 6 and Table 1"},{"comment":"The heuristic that selects the kernel and length-scale by minimizing the condition number of the MMD Hessian is evaluated on the same problem instances used for the final comparison; it would be helpful to state whether this selection is performed on a separate validation run or is part of the reported algorithm, to avoid any appearance of tuning on the test set.","section":"Section C.2"},{"comment":"In Theorem 1, the definition of λ as a ratio of integrals requires the denominator integral to be nonzero; the proof should discuss when this holds and what happens in the degenerate case, especially since Example 2 then approximates the ratio by m2/m1.","section":"Section A.2 and Example 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and the empirical results are suggestive, but the two main advertised theoretical/empirical claims both need significant strengthening. The Lemma 1 issue is a genuine mathematical gap that should be fixed or demoted to a sufficient condition. More importantly, the reference-set construction makes the central 'imperfect reference' robustness claim nearly unfalsifiable in the current protocol: R is always derived from Y0, so the one loss and the Section 7 concession are consistent with the circularity. I would advise the editor to send the paper back for major revision rather than reject, because the derivations and the hybrid framework could be salvageable with a clearer separation between local refinement and coverage recovery, plus ablations that decouple R from Y0."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The genuinely new thing is the analytical gradient and Hessian of MMD^2 with respect to the decision vectors, and the Newton method built on them. That part is real and, as far as I can check, correct. The chain-rule derivations in Section 4.1 are standard, the eigenspectrum bounds in Theorem 2 are plausible, and the hybrid with MOEAs is a reasonable idea. The experimental setup is also fair-minded: they account for the cost of Jacobian/Hessian evaluations via measured AD times, and the 28/33 wins over three baselines is consistent evidence that the post-processing step helps on these benchmarks. Improvements are modest in absolute terms, but real.\n\nThe soft spots, in proportion. First, Lemma 1 claims an `if and only if` stationary condition, but the proof jumps from `integral of omega times bracket equals zero` to `bracket vanishes pointwise` because G is non-degenerate. That inference is invalid: a non-negative measure can integrate a signed function to zero without pointwise vanishing. The `if` direction is fine; the `only if` is not established. This is a theoretical gap, not a fatal one, but it should be fixed.\n\nSecond, and more important, the central robustness claim -- that MMD's diversity term rescues an imperfect reference set -- is not supported by the experiments. The reference set R is built from the MOEA's own output Y0 by interpolation, k-means, and a shift toward the utopian region. So R is a smoothed, shifted copy of the very set being refined. When Y0 covers only a subregion, R inherits that bias, and the MMD cross-term pulls points toward R's support while the self-term spreads them within that support, not onto the missing part. The one loss in Table 1, MOEA/D on DTLZ7, is exactly this failure, and Section 7 concedes it. In the 28 wins, Y0 plausibly already spans the front, so the improvement may be ordinary local Newton refinement of a good warm start rather than the diversity mechanism. The only direct evidence for recovery is Fig. 1, a single unquantified illustration. I think the paper should either temper the abstract's claim or test recovery with deliberately biased initial populations.\n\nMinor things: the code link points to a HypervolumeDerivatives repository, which looks mismatched; and kernel/length-scale are chosen per problem by minimizing the condition number, which is tuning on the benchmark. Both should be cleaned up.\n\nOverall, the method is useful as a post-processing tool for differentiable multi-objective problems, and the analytical derivatives are worth having. The theory and the reference-set attribution both need work, but the paper deserves a serious referee and likely major revision.","headline":"MMD-Newton is a real but modest contribution: the analytical derivatives are correct and the hybrid idea is sensible, but the reference-set robustness story is oversold and Lemma 1 has a gap.","tokens_in":22056,"tokens_out":1947,"would_cite":false,"duration_ms":21222,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C29","90C53"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes minimizing the maximum mean discrepancy between a Pareto approximation set and a reference set, and shows that a Newton method on that objective both pulls the set onto the front and keeps it spread, beating three…","keywords":["multi-objective optimization","maximum mean discrepancy","Newton method","Pareto front approximation","kernel mean embedding","evolutionary algorithm","reference set","Hessian eigenspectrum"],"falsifier":"Take a two- or three-objective problem whose Pareto front has two well-separated components; run the warm-start MOEA for 300 iterations and stop it before any population member enters the second component; build the reference set from that population and run MMDN for five iterations; if the final approximation still has no points on the second component, the claimed ability to recover from an imperfect reference is contradicted in that setting.","tokens_in":20865,"feed_emoji":"🎯","tokens_out":8218,"duration_ms":73114,"temperature":0.7,"pith_summary":"The paper tries to establish that maximum mean discrepancy (MMD) is a suitable distance between a Pareto approximation set and a reference set for continuous multi-objective optimization, and that minimizing it with a second-order method refines fronts faster than evolutionary search alone. Treating the two finite sets as empirical measures, the authors derive the gradient and Hessian of $\\mathrm{MMD}^2$ with respect to the decision variables and build a Newton method, called MMDN, with theoretical conditions on stationarity and on the Hessian's eigenvalues. Since squared MMD contains the RKHS norm of the approximation set's embedding, minimizing it also spreads the set, so refinement works even when the reference does not cover the whole front. Hybridized with an MOEA, which supplies a warm start and a reference, MMDN produced better average Hausdorff distances than the MOEA alone under the same evaluation budget in 28 of 33 benchmark configurations.","feed_headline":"MMD-Newton sharpens Pareto fronts with imperfect references","feed_subtitle":"The metric pulls points onto the front while spreading them, beating MOEAs on 28 of 33 benchmark problems.","key_machinery":"The central object is squared maximum mean discrepancy between two empirical measures, $$\\mathrm{MMD}^2(Y,R)=\\|\\psi(Y)-\\psi(R)\\|_{\\mathcal H}^2=\\frac{1}{\\$mu^{2}$}\\sum_{i,j}k(y_i,y_j)+\\frac{1}{\\$lambda^{2}$}\\sum_{i,j}k(r_i,r_j)-\\frac{2}{\\mu\\$\\lambda$}\\sum_{i,j}k(r_i,y_j),$$ where $\\psi$ is the kernel mean embedding into an RKHS with a Gaussian or Matérn kernel. The carrying mechanism is the Newton step for the KKT system of minimizing $\\mathrm{MMD}^2$ over stacked decision variables $X\\in\\mathbb{R}^{\\mu n}$ subject to constraints, using the derived gradient and Hessian and an active-set treatment of inequalities. Because the RKHS norm $\\|\\psi(Y)\\|_{\\mathcal H}^2$ grows when points move apart under a decreasing kernel, the method spreads $Y$ while pulling it toward $R$; because the Hessian is not guaranteed positive definite, the algorithm preconditions it and uses a backtracking line search.","core_discovery":"On the paper's own terms, the core discovery is that a Newton method on $\\mathrm{MMD}^2(Y,R)$ can refine a Pareto approximation set $Y=F[X]$ against a reference set $R$, and the diversity term hidden inside MMD is what makes the refinement robust to an imperfect $R$. The authors provide closed-form first and second derivatives of $\\mathrm{MMD}^2$ with respect to each decision point, characterize the first-order stationary condition in the frequency domain, and bound the Hessian spectrum, showing it can be indefinite and therefore must be preconditioned. Empirically, warm-starting MMDN with the output of NSGA-II, NSGA-III, or MOEA/D and running five Newton steps improves the approximation accuracy compared with running the MOEA alone under the same function-evaluation budget, with one clear failure on DTLZ7 when the warm-start population is concentrated on a subregion.","pith_inferences":["If the diversity mechanism is as general as the Figure 1 example suggests, the same MMD minimization could serve as a post-processing step for any set-generating optimizer, including Bayesian or gradient-based multi-objective methods, not just MOEAs.","The stationary-condition analysis predicts that the converged location of each point is tied to the kernel length-scale and to the reference set's center of mass, so choosing $\\theta$ tunes a trade-off between fitting the reference and maximizing spread.","A natural stress test is disconnected Pareto fronts: MMD should preserve coverage only if the reference or the initial set samples every component, and the paper's failure case on DTLZ7 suggests the recovery claim has a coverage threshold."],"forward_implications":["MMDN can be used as a cheap post-processor: five Newton steps after a 300-iteration MOEA run improve the Pareto approximation under the same total evaluation budget.","The diversity term in MMD means that a reference set that does not cover the full front does not necessarily cap the final coverage; the method can spread beyond it.","The Hessian spectral analysis explains why preconditioning is necessary and identifies the spread-dependent terms that control the spectrum.","The hybrid's accuracy is not uniform: on problems where the warm-start population is concentrated, such as DTLZ7 with MOEA/D, the refinement can be worse than the MOEA alone."],"supporting_citations":[{"why":"defines MMD and the kernel two-sample statistic that the paper minimizes.","marker":"[17]"},{"why":"gives the kernel mean embedding and Bochner spectral representation used to derive gradients, Hessians, and spectrum bounds.","marker":"[25]"},{"why":"supplies the constrained Newton framework, active-set treatment, Hessian preconditioning, and backtracking line search.","marker":"[26]"},{"why":"provides the Newton-style set refinement and the reference-set construction procedure that MMDN adapts.","marker":"[35]"},{"why":"provides the reference-set generation operations of filling, reduction, and shifting used in Algorithm 1.","marker":"[27]"},{"why":"supplies NSGA-II, one of the three warm-start baselines in the empirical comparison.","marker":"[8]"},{"why":"supplies NSGA-III, one of the three warm-start baselines in the empirical comparison.","marker":"[9]"},{"why":"supplies MOEA/D, one of the three warm-start baselines and the case where the hybrid loses on DTLZ7.","marker":"[38]"},{"why":"defines the ZDT benchmark problems used in the experiments.","marker":"[39]"},{"why":"defines the DTLZ benchmark problems used in the experiments.","marker":"[10]"}],"fun_headline_variants":["MMD-Newton: closed-form Hessian refines Pareto fronts","Newton on MMD beats MOEAs with warm start","Pareto refinement via MMD-Newton with imperfect references","A Newton method for multi-objective optimization","MMD-Newton: hybrid with MOEA refines Pareto sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success depends on the reference set built from the MOEA's own starting population; if that population covers only a piece of the Pareto front, the reference inherits the gap and the refinement cannot reliably recover the missing piece.","fun_headline_variants_meta":{"raw":{"variants":["MMD-Newton: closed-form Hessian refines Pareto fronts","Newton on MMD beats MOEAs with warm start","Pareto refinement via MMD-Newton with imperfect references","A Newton method for multi-objective optimization","MMD-Newton: hybrid with MOEA refines Pareto sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3561,"prompt_tokens":991,"completion_tokens":2570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2487}},"tokens_in":607,"tokens_out":2570,"duration_ms":19164,"temperature":1.0,"reasoning_tokens":2487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:31:13.597788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two- or three-objective problem whose Pareto front has two well-separated components; run the warm-start MOEA for 300 iterations and stop it before any population member enters the second component; build the reference set from that population and run MMDN for five iterations; if the final approximation still has no points on the second component, the claimed ability to recover from an imperfect reference is contradicted in that setting.","supporting_citations":[{"cited_title":"Borgwardt, Malte J","cited_arxiv_id":null,"evidence_quote":"defines MMD and the kernel two-sample statistic that the paper minimizes."},{"cited_title":"Sriperumbudur, and Bernhard Schölkopf","cited_arxiv_id":null,"evidence_quote":"gives the kernel mean embedding and Bochner spectral representation used to derive gradients, Hessians, and spectrum bounds."},{"cited_title":"Wright.Numerical Optimization","cited_arxiv_id":null,"evidence_quote":"supplies the constrained Newton framework, active-set treatment, Hessian preconditioning, and backtracking line search."},{"cited_title":"A Newton Method for Hausdorff Approximations of the Pareto Front within Multi-objective Evolutionary Algorithms","cited_arxiv_id":"2405.05721","evidence_quote":"provides the Newton-style set refinement and the reference-set construction procedure that MMDN adapts."},{"cited_title":"Rodriguez-Fernandez, Hao Wang, and Oliver Schütze","cited_arxiv_id":null,"evidence_quote":"provides the reference-set generation operations of filling, reduction, and shifting used in Algorithm 1."},{"cited_title":"Meyarivan","cited_arxiv_id":null,"evidence_quote":"supplies NSGA-II, one of the three warm-start baselines in the empirical comparison."},{"cited_title":"An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"supplies NSGA-III, one of the three warm-start baselines in the empirical comparison."},{"cited_title":"MOEA/D: A multiobjective evolutionary algorithm based on decomposition.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"supplies MOEA/D, one of the three warm-start baselines and the case where the hybrid loses on DTLZ7."},{"cited_title":"Comparison of multiobjective evolutionary algorithms: Empirical results.Evolutionary computation, 8(2):173–195, 2000","cited_arxiv_id":null,"evidence_quote":"defines the ZDT benchmark problems used in the experiments."},{"cited_title":"Scalable Test Prob- lems for Evolutionary Multiobjective Optimization","cited_arxiv_id":null,"evidence_quote":"defines the DTLZ benchmark problems used in the experiments."}],"review_version":1}