{"id":"4e260dbf-bf93-463f-bb55-8d33894fb44f","arxiv_id":"2510.14122","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MiSSOC solves challenging MINLPs by fitting shape-constrained additive B-spline surrogates to complicating functions and solving the resulting separable surrogate with a tailored solver.","lead":"MiSSOC is a new algorithm that tackles hard mixed-integer nonlinear optimization problems by replacing messy nonlinear pieces with smooth spline approximations learned from sample points, optionally obeying expert rules about shape, and then solving this simpler stand-in problem to guide a local search. The paper is worth reading because it offers a practical route to problems that commercial solvers time out on, with encouraging but limited tests on benchmarks and a real hyd","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MiSSOC's competitiveness claim is not yet established: the experiments compare against 600-second global-solver runs but omit a random multistart + local-search control, so the surrogate may not be the component producing the reported solutions.","rationale":"The paper is honest about its main theoretical limitation: Section 1 explicitly states that no error bound is guaranteed, and Section 4.1 lists why approximation-based surrogates are heuristic. The numerical work is internally coherent, and the public code is a real asset. My concern is not that the additive assumption is false in every instance; it is that the experiments as designed cannot establish the central competitiveness claim. The final local search can improve any starting point, so observed values are a joint product of the surrogate and IPOPT. The MINLPlib results actually undercut 'competitive': direct solvers recover the same or better best-known values in all three instances, and on ex6_2_5 MiSSOC is 0.27% worse while direct solvers find the optimum. The HUC result is a single real case with a tiny improvement over the published value, and it depends on solver and degree. A random multistart control is cheap and would determine whether the surrogate is doing the work. If it is not, the paper should be read as presenting a valid surrogate pipeline, not as proving MiSSOC is competitive. This is consistent with the reader's CONDITIONAL verdict: I do not see grounds to reject the method, only to reject the current strength of the claim. The reader's additivity concern is real but secondary for the empirical claim; the missing multistart baseline tests whether the surrogate contributes at all. Thus the verdict remains conditional pending that baseline.","tokens_in":24084,"tokens_out":8846,"duration_ms":82025,"concrete_test":"Run a random multistart + local-search baseline on the same instances with the same total time budget as MiSSOC (sampling, fitting, surrogate solve, and local search). For each instance, draw 10 seeds times 100-1000 random integer assignments (for HUC, random commitment patterns; for MINLPlib, random feasible starts in the box, discarding infeasible ones), fix the integers, solve the resulting NLP with IPOPT under a per-run time budget, and record best g0, mean, and time-to-best. If the baseline matches or beats the values in Tables 2 and 4, the surrogate is not load-bearing and the 'competitive algorithm' claim should be downgraded; if it does not, the claim is supported by this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's final step fixes the integer variables from the surrogate solution and runs a local NLP solve. The reported g0(x*) therefore reflects the surrogate plus IPOPT polish, not the surrogate alone. The paper never compares against a random multistart baseline: sample many integer assignments, fix them, run the same NLP local solver, and take the best. Without this control, the experiments cannot separate the contribution of the additive B-spline surrogate from the contribution of local search or instance structure. This gap is load-bearing because the abstract and Section 6 assert that MiSSOC 'is a competitive algorithm,' not merely that a surrogate pipeline exists. The MINLPlib results in Table 2 further weaken the claim: direct global solvers already find the best-known values for ex6_2_13 and ex6_2_7 and a strictly better value (-70.752) than MiSSOC's (-70.558) on ex6_2_5, so the advantage is time-to-solution, not solution quality. On HUC, the reported improvement over the published 14533.1 is only 14537.257 (Table 4, degrees 5-7), a 0.03% gap, and no repeated trials are reported. The additivity limitation of Section 3 is a related theoretical risk, but the missing multistart control is the directly testable weakness that currently blocks the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MiSSOC, a surrogate-based algorithm for nonconvex MINLPs. Complicating functions (initially the objective) are approximated by constrained smooth additive B-spline regression models; the resulting surrogate is cast as an MINLP with a multiple-choice formulation and solved with SC-MINLP or general-purpose global solvers, followed by a local NLP polish on the original problem. The methodological contribution includes a conic formulation for shape-constrained additive regression. Experiments are reported on three MINLPlib instances and a Hydro Unit Commitment case study, with comparisons against direct global-solver runs. The central claim is that MiSSOC, especially with SC-MINLP, is a competitive algorithm for challenging MINLPs.","tokens_in":24472,"tokens_out":4141,"duration_ms":39070,"significance":"If the empirical claims are established, the paper offers a useful and original integration of statistical shape-constrained regression into MINLP surrogate construction. The use of additive separable spline surrogates is well matched to the SC-MINLP solver, and the public code/data repositories are a strength. The methodological extension from univariate to additive shape constraints, while sufficient-only for the sum constraint, is a plausible contribution. However, the experimental section as written does not yet support the abstract's and Section 6's competitiveness claim: the reported advantages are not isolated from the effect of the final local search, the optimality gap is reported for the surrogate rather than the original problem, and the randomness of the sampling phase is not controlled. These are fixable with additional experiments and careful reporting.","major_comments":[{"comment":"The central competitiveness claim is not supported without a random-multistart + local-search control. Algorithm 1's final step fixes the surrogate's integer decisions and solves an NLP with IPOPT; on the three MINLPlib instances there are no integer variables at all, so the pipeline is surrogate-selection plus local NLP polish. The paper compares only to 600-second runs of global solvers, not to a simple baseline that samples many starting points/integer assignments and applies the same IPOPT local search. Such a baseline is needed to separate the contribution of the additive B-spline surrogate from the contribution of local search and instance structure. The concern is concrete: in Table 2, MiSSOC's value on ex6_2_5 (-70.558) is worse than the MINLPlib best (-70.752) that BARON and COUENNE find directly, so the only reported advantage on that instance is time-to-solution. Without the c","section":"§5, Algorithm 1 and Tables 2/4"},{"comment":"The optimality gap reported for the surrogate problem is not a gap for the original MINLP. Equation (23) is explicitly 'computed in the solution of the surrogate problem (P~), that is, before the local search phase.' A zero or near-zero gap in Tables 2 and 4 only shows that the fitted surrogate was solved to optimality. Since Section 1 acknowledges that MiSSOC 'cannot guarantee an error bound over the entire domain,' a zero surrogate gap can coexist with substantial suboptimality or infeasibility in the original problem. The narrative in §5.1 and §5.2 repeatedly uses the small surrogate gaps as evidence of solution quality. The paper should report a gap or bound for the original problem (e.g., against a valid dual bound or best-known value), or explicitly and consistently label the reported gap as a surrogate-solving metric only.","section":"§5, Eq. (23)"},{"comment":"The experiments use one run per configuration despite random training-data sampling. Section 4.3 specifies that the training set T is generated by uniform random sampling, and Section 5.1 states '15 samples per parameter' are drawn. None of the reported times, objective values, or gaps are accompanied by repeated seeds, standard deviations, or median/min-max summaries. This is important because the training sample directly determines the fitted surrogate and hence the final solution. For example, the near-instantaneous SC-MINLP times (0.028s in Table 2) and the degree-dependent HUC results in Table 4 may be particular to one draw. The authors should either report repeated runs with different seeds or justify that the results are stable across sampling.","section":"§5.1 and §5.2"}],"minor_comments":[{"comment":"The text says 'Constraints (21), (21) and (22) stated the nature of all the variables.' The duplicate '(21)' should be '(21), (22)' or similar, and 'stated' should be 'state.'","section":"§4.2, constraints (21)-(22)"},{"comment":"There is a typo: 'More specifically, (Bj)il = Bl,dj,tj (xij), for i = 1, . . . , n and l = 1, . . . , kk + dj' — 'kk' should be 'kj.'","section":"§2.3"},{"comment":"The phrase 'proves that MiSSOC is a competitive algorithm' is too strong given the experimental limitations discussed above. Recommend 'indicates' or 'provides evidence that.'","section":"§6"},{"comment":"The claim that degrees 5-7 'improve' the published value 14533.1 should be stated with the sense of the objective (maximization) explicitly in the surrounding text, since the same table is used for the minimization MINLPlib instances. This is clear from context but should be made explicit to avoid confusion.","section":"§5.2, Table 4"},{"comment":"The rule '15 samples per parameter' is introduced without justification or sensitivity analysis. A brief comment on why 15 is sufficient, or a small sensitivity study, would strengthen the sampling discussion.","section":"§5, sampling size rule"}],"recommendation":"major_revision","confidential_remarks":"The methodological idea is sound and the paper is well written, but the experimental validation currently does not support the headline claim. The missing random-multistart control is the key blocker; it is directly addressable within the manuscript's scope. I would not reject, because the surrogate-plus-SC-MINLP pipeline may well be competitive once properly benchmarked. Please ensure the revised version also clarifies the distinction between surrogate gaps and original-problem quality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: MiSSOC is a competent, honest paper. The core idea — approximate the hard functions of an MINLP with shape-constrained additive B-splines, then solve the resulting separable surrogate with a tailored MINLP solver — is a genuine combination. The additive extension of Navarro-García et al.'s univariate shape-constrained B-spline regression (Section 3) is new as far as I know, and the weight-based decomposition of bound constraints (Eq. 13) is a reasonable workaround. The multiple-choice MINLP formulation in Section 4.2 is standard but well suited to the separable structure, and the SC-MINLP integration is a sensible use of available machinery. Code appears to be available and the fitting steps are reproducible.\n\nThe experiments tell a weaker story than the abstract and Section 6 claim. The three MINLPlib instances are small and continuous, and direct solvers already reach the best-known solutions for two of them and a better value on ex6_2_5. The reported advantage is time-to-solution, and the 'proves' language in the conclusions oversells it. On HUC, the improvement over the published value is about 4 units on ~14533 (0.03%), with no repeated trials. More importantly, the final step of Algorithm 1 fixes the integer variables from the surrogate and runs IPOPT; without a random multistart + local-search baseline, you cannot tell whether the surrogate or the local polish is doing the work. That is a load-bearing gap for the 'competitive algorithm' claim, and the stress-test note has it right.\n\nThe additivity assumption is a real limitation, but the authors acknowledge it: no error bound over the domain, and they explicitly say the approach is heuristic. I don't think that is a red flag; it's the nature of the method. The circularity concern about reporting the surrogate's zero optimality gap (Eq. 23) as evidence of solution quality is fair — that gap measures the surrogate, not the original — but they do report g0(x*) for all solutions, so it's more a presentation issue than a flaw.\n\nWho should read this: people working on surrogate-based MINLP or shape-constrained regression. It deserves a serious referee: the method is reproducible, the writing is clear, and the missing baseline is fixable with a modest computational study. I would ask for random multistart comparisons, repeated seeds, and a softening of 'proves' to 'suggests'.","headline":"MiSSOC is a sensible surrogate-MINLP pipeline with a real methodological addition, but the experiments don't yet isolate what the surrogate actually contributes.","tokens_in":24973,"tokens_out":3440,"would_cite":true,"duration_ms":29112,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C26","62G08","65D07"],"pacs":[],"model":"deepseek-v4-flash","headline":"MiSSOC solves challenging mixed-integer nonlinear programs by replacing complicated functions with shape-constrained sums of B-spline curves, then exploits the resulting separable structure to find high-quality solutions fast.","keywords":["mixed-integer nonlinear programming","surrogate modeling","shape-constrained regression","B-splines","additive regression","separable nonconvex optimization","global optimization","hydro unit commitment"],"falsifier":"Take a test MINLP whose objective contains a strong interaction term, such as x1*x2*x3, fit the additive B-spline surrogate, and solve it. Then compare the true objective value at the surrogate's solution with the best value found by a reliable global solver on the original problem: if the surrogate's solution is consistently much worse while its in-sample fit is excellent, the additivity assumption is the failure point.","tokens_in":24050,"feed_emoji":"🧮","tokens_out":3530,"duration_ms":35936,"temperature":0.7,"pith_summary":"The paper introduces MiSSOC, an algorithm that builds a tractable surrogate of a hard mixed-integer nonlinear program (MINLP) by approximating each complicating function with a smooth additive regression model made of univariate B-spline curves. Expert knowledge about bounds, monotonicity, or curvature can be imposed on these curves, so the surrogate is both data-driven and knowledge-driven. The surrogate keeps simple nonlinearities intact but turns nonconvex complications into sums of one-variable polynomials, a form that specialized solvers can handle efficiently. The authors claim that MiSSOC, especially when paired with the SC-MINLP solver, finds high-quality solutions for challenging MINLPs in far less time than solving the original problem directly.","feed_headline":"Surrogate splines solve hard MINLPs in under a second","feed_subtitle":"Replacing hard nonlinear functions with shape-constrained B-spline sums makes mixed-integer global optimization tractable.","key_machinery":"The central object is the smooth additive regression model: each complicated function g is approximated as a constant plus a sum of univariate B-spline functions over the box defined by variable bounds. B-splines are piecewise polynomials joined at knots, so the fitted function is a sum of univariate piecewise polynomials. Expert knowledge enters through conic positive-semidefinite constraints, adapted from univariate shape-constrained smoothing, that enforce bounds, monotonicity, or curvature on each additive component. The surrogate is then written as a mixed-integer program using a multiple-choice formulation, where each knot interval contributes a binary variable; this formulation has se","core_discovery":"The central claim is that a nonconvex MINLP, whose bottleneck is a complicated objective or constraint function, can be solved effectively by first sampling that function, fitting a smooth additive B-spline approximation with optional shape constraints, and then solving the resulting surrogate as an MINLP with separable structure. The paper argues that this surrogate is accurate enough that its optimal solution, after a local-search refinement, matches or improves solutions found by general-purpose global solvers on the original problem. In the reported experiments, solving the surrogate with the SC-MINLP algorithm yields near-zero optimality gaps in fractions of a second, while direct solve","pith_inferences":["Beyond the paper: the additive form limits MiSSOC to problems where interactions between variables are weak; for interaction-heavy functions, one would expect the surrogate to miss ridges and valleys, so a natural extension is adding bivariate spline terms or an interaction-detection pre-pass.","Beyond the paper: uniform random sampling may be wasteful; an adaptive sampling strategy that concentrates points near promising regions, as the paper lists as future work, would likely improve both approximation quality and final solution value.","Beyond the paper: the weights that allocate the global bound among additive components are currently fixed by data; treating them as decision variables in the fitting problem could tighten the bounds and make the surrogate more faithful.","Beyond the paper: if shape constraints were chosen so the surrogate under- or overestimates the original function, the surrogate would become a certified relaxation or restriction of the original MINLP, turning a heuristic into a method with provable bounds."],"forward_implications":["If the central claim holds, hard nonconvex MINLPs with complex objectives can be approached by sampling and spline fitting instead of direct global optimization, often with dramatic time savings.","Combining surrogate building with SC-MINLP can produce zero or near-zero optimality gaps on the surrogate, meaning the separable structure of B-spline sums is the key to tractability.","Shape constraints allow practitioners to inject domain knowledge—such as known output bounds or monotonic behavior—into the surrogate, so the optimization respects qualitative physical or economic facts.","The same machinery extends to approximating constraint functions, opening a route to surrogate-based feasibility recovery for problems where constraints, not just the objective, are the complication.","The reported improvement over the best known solution of a real hydroelectric scheduling instance suggests the method can yield practically better decisions, not just faster ones."],"fun_headline_variants":["Spline surrogates crack MINLPs in milliseconds","Shape-constrained splines tame tough MINLPs quickly","MISSOC: B-spline surrogates for fast global optimization","Spline surrogates hit near-zero gaps in under a second"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole approach depends on the complicated function being well represented as a sum of univariate B-spline curves across the variable bounds, so that the surrogate's optimum tracks the true optimum; the paper gives no error bound to certify this approximation.","fun_headline_variants_meta":{"raw":{"variants":["Spline surrogates crack MINLPs in milliseconds","Shape-constrained splines tame tough MINLPs quickly","MISSOC: B-spline surrogates for fast global optimization","Spline surrogates hit near-zero gaps in under a second"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00037,"raw_usage":{"total_tokens":1858,"prompt_tokens":819,"completion_tokens":1039,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":967}},"tokens_in":563,"tokens_out":1039,"duration_ms":8964,"temperature":1.0,"reasoning_tokens":967,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:36:36.076853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test MINLP whose objective contains a strong interaction term, such as x1*x2*x3, fit the additive B-spline surrogate, and solve it. Then compare the true objective value at the surrogate's solution with the best value found by a reliable global solver on the original problem: if the surrogate's solution is consistently much worse while its in-sample fit is excellent, the additivity assumption is the failure point.","supporting_citations":[],"review_version":1}