{"id":"3204c54d-73c9-4a5f-a135-a95f268d8b00","arxiv_id":"1908.02149","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A deterministic fractal-decomposition optimizer, scalarized with Tchebycheff weights and parallelized via containers, ranks first among six algorithms on three of four benchmark quality metrics.","lead":"This paper combines a fractal-based search algorithm with a standard trick for multi-goal problems and runs many copies in parallel containers across machines. It claims this approach beats well-known algorithms on eight standard test problems and runs faster on two machines than on one.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mo-FDA's claimed empirical superiority is unauditable: max evaluations, n, and run counts are missing, k is tuned without disclosure, and Table 1 shows a Friedman-rank arithmetic anomaly.","rationale":"Agree with the reader: the load-bearing weakness is the auditability of the empirical comparison. Mo-FDA's advantage is asserted through a Friedman-rank table, but the evaluation budget, n, number of runs, and the search over k are undisclosed. Since the algorithm is deterministic and the baselines are stochastic, a fair comparison requires at least matched budgets and repeated runs; without these, the rank differences in Table 1 could be an artifact. The additional arithmetic observation (Hypervolume row sums to 20.75 rather than 21) reinforces the need for raw data. This is not a reason to reject the approach outright: the algorithm description is coherent, the container architecture is plausible, and the claimed timing pair (0.8s to 0.5s) could be true. The issue is that the paper does not supply enough information to test the central empirical claim. A conditional acceptance requiring full settings, raw per-problem values, and statistics is appropriate; hence verdict unchanged from the reader.","tokens_in":3990,"tokens_out":10369,"duration_ms":109151,"concrete_test":"Re-run the benchmark with an explicit protocol: fix a single evaluation budget for Mo-FDA and all five baselines (e.g., the budget from Jiang et al. [3]), declare n, run the stochastic algorithms at least 30 times, and recompute the per-metric Friedman ranks from the raw values. As part of the same check, verify that each row of average ranks sums to 21; the Hypervolume row currently sums to 20.75 and must be corrected or explained. If Mo-FDA still ranks first on Hypervolume, GD, and Spread under this audit, the claim stands; if not, it fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Mo-FDA ranks first on Hypervolume, GD, and Spread over eight ZDT/DTLZ problems (Table 1). The only supporting evidence is the Friedman-rank table, but the protocol generating it is not stated. Section 4 says only that 'the principal experiments settings described in [3]' were used; Section 2 identifies k=5 as 'experimentally determined' without saying how or on which problems. The maximum number of evaluations that terminates the exploitation phase and the number of weight combinations n (which determines the Pareto-front size and the number of containers) are never reported. Because Mo-FDA is deterministic and its cost and solution quality depend directly on these quantities, any mismatch with the jMetal 5.0 baselines' budgets would make the comparison unfair in an unknown direction. No per-problem metric values, number of stochastic runs, or significance tests are given, so the margins in Table 1 cannot be checked. The table itself is also arithmetically suspicious: for 6 algorithms and 8 problems the average Friedman ranks per metric should sum to 21, but the Hypervolume row sums to 20.75, implying a misreported value or an omitted algorithm/problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Mo-FDA, a deterministic multiobjective extension of the Fractal Decomposition Algorithm. The algorithm scalarizes the multiobjective problem with a weighted Tchebycheff function and solves n independent scalarized instances, one per weight vector, using a container-based parallel multi-node strategy. The authors benchmark Mo-FDA against NSGA-II, NSGA-III, MOEA/D, GWASFGA, and CDG on five ZDT and three DTLZ problems, reporting average Friedman ranks (Table 1) in which Mo-FDA ranks first on Hypervolume, GD, and Spread, and second on IGD. They also report that the container architecture reduces the average computation time from 0.8 seconds on one host to 0.5 seconds on two hosts.","tokens_in":4151,"tokens_out":5257,"duration_ms":51580,"significance":"If the reported results were fully reproducible, the contribution would be useful: a deterministic alternative to stochastic evolutionary MO solvers with a portable containerized parallelization is a legitimate practical advance. The algorithmic construction is coherent: the scalarization is standard and the comparison with external baselines on public benchmarks avoids circularity. However, the paper currently provides insufficient experimental detail to verify the central performance claim, and Table 1 contains an arithmetic anomaly. With a complete protocol and corrected statistics, the contribution would be a solid short empirical paper; as written, the evidence is not yet auditable.","major_comments":[{"comment":"The Hypervolume row of Table 1 is arithmetically inconsistent: the six reported average Friedman ranks sum to 20.75 instead of the expected 21.0 for six algorithms over eight problems, meaning either a value is misreported or one problem/algorithm is omitted. Because the paper's central claim relies on this table, please correct the table and provide the underlying per-problem metric values so the ranks can be verified.","section":"Section 4, Table 1"},{"comment":"The manuscript states that 'the principal experiments settings described in [3]' were used, but it does not enumerate those settings. In particular, the maximum number of evaluations, the number of weight combinations n, the population sizes and termination criteria for the jMetal 5.0 baselines, and the number of independent runs for stochastic algorithms are never reported. Since Mo-FDA is deterministic and its solution quality and runtime depend directly on n and the evaluation budget, the comparison cannot be audited and may be unfair in an unknown direction. Please report all experimental settings explicitly in the paper.","section":"Section 4, experimental protocol"},{"comment":"The sentence 'k has been experimentally determined and set to 5' gives no information about the tuning process. If k was tuned on the same benchmark problems used in the final comparison, the reported performance of Mo-FDA could reflect overfitting. Please describe the tuning procedure, the range of k tested, the problems on which k was chosen, and confirm that k was fixed before benchmarking.","section":"Section 2, fractal depth k"},{"comment":"Table 1 reports only average Friedman ranks, without variance, number of runs, or any significance test. The comparison involves stochastic baselines, so the reported margins (e.g., 1.875 vs 2.25 in Hypervolume) may not be meaningful. Report the number of independent runs per algorithm, the standard deviations of the metrics, and the results of a suitable significance test (e.g., Wilcoxon signed-rank with Holm correction) over the eight problems.","section":"Section 4, statistical reporting"},{"comment":"The claimed speedup from 0.8 seconds on one host to 0.5 seconds on two hosts is not reproducible as stated. The manuscript does not specify the number of weight combinations n, the number of containers per host, whether the measured time is total wall-clock time including container startup and aggregation, or which of the eight functions and which dimension (D=30 is mentioned but the sentence is garbled) are involved. Because the multi-node container strategy is a central contribution, please define the measurement precisely and report the configuration used.","section":"Section 4, time measurement"}],"minor_comments":[{"comment":"The sentence 'The rest of the of paper is organized as follow' contains a typo and a subject-verb disagreement; it should read 'The rest of the paper is organized as follows.'","section":"Introduction"},{"comment":"The algorithm names are used inconsistently: the text and table use 'MEOA/D' and 'MEOAD/D', while the standard abbreviation is 'MOEA/D'; similarly the table header contains 'GW ASGF A' with an extra space. Please standardize the nomenclature.","section":"Section 3 and Table 1"},{"comment":"The caption 'Functions where Mo-FDA performs the best and the worst respectively' is ambiguous because the figure shows two plots but the caption does not state which plot corresponds to 'best' and which to 'worst'. Clarify the caption and improve the figure resolution and axis labels for readability.","section":"Figure 1"},{"comment":"The paper states that the sum of weights ωi must equal 1 but does not describe how the n weight vectors are generated (e.g., uniform grid, random sampling, or a specific decomposition method). This affects the coverage and spread of the resulting Pareto-front approximation and should be specified.","section":"Section 2, weight vectors"},{"comment":"It is not explained how the n scalarized solutions are combined into the final Pareto-front, for example whether dominated solutions are filtered out and how ties or duplicate points are handled. This is needed to interpret the metric values in Table 1.","section":"Section 2, Pareto-front construction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short empirical report and the main issue is reproducibility rather than the algorithmic idea. The editor may wish to require supplementary material with full experimental configurations and raw metric values, and to ask the authors to fix the arithmetic inconsistency in Table 1, before further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a short empirical paper, not a theory paper. The new bit is taking the authors' deterministic FDA optimizer, scalarizing with Tchebycheff weights, and running n weight-vector instances in separate containers on one or two hosts. That combination is genuinely new as far as I know, and the container idea is a reasonable way to get parallel speedup without touching the algorithm's internals. I believe the central mechanism works as described: each container solves one scalarized problem, and the Pareto front is assembled from the solutions.\n\nThe fair reading is that the paper gives Mo-FDA credit for being deterministic, simple, and parallelizable, and the abstract is honest about scope. That is worth something.\n\nNow the soft spots. The empirical comparison is under-specified to the point that the headline claim, 'best Friedman rank on three of four metrics', cannot be audited. The authors say they borrowed 'the principal experiments settings' from Jiang et al. [3] but do not tell the reader which settings those are. The maximum evaluation budget, the number of weight vectors n, and the number of independent runs per algorithm are all missing. The fractal depth k is said to be 'experimentally determined' and set to 5, with no indication of the tuning procedure or the data used for tuning. Since Mo-FDA's cost and output quality depend directly on k and the budget, this is a real gap, not a minor omission. The table itself has an arithmetic anomaly: for six algorithms and eight problems, the average Friedman ranks per metric must sum to 21, but the Hypervolume row sums to 20.75. That suggests a typo, but it makes the reader suspect the table was assembled by hand.\n\nAlso missing: per-problem metric values, error bars, significance tests. With a deterministic algorithm one might compare against deterministic baselines, but the jMetal baselines are stochastic, so variability matters. The scalability evidence is a single timing pair: 0.8 s on one host, 0.5 s on two. That is not enough to establish scalability.\n\nNone of this is fatal in the sense of a wrong derivation; it is a reporting problem. The methods are reproducible in principle, and the missing details are fixable. As written, though, the evidence for the central claim is not there.\n\nWho benefits: someone working in decomposition-based multiobjective optimization who wants a quick read on whether a deterministic FDA with containers is worth trying. The citation pattern is appropriate; self-citation of the FDA paper is justified.\n\nRecommendation: I would give it a referee cycle at a workshop or a short-paper track, but not a full journal. If it comes back with full experimental settings, code, and the table fixed, it could be a decent contribution. As is, it is a preliminary report.","headline":"Plausible deterministic multiobjective scheme, but the empirical case as written is unauditable: missing settings, tuned k, no statistics, and a Friedman-row sum that does not add up.","tokens_in":4711,"tokens_out":3017,"would_cite":false,"duration_ms":29130,"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":"This paper claims that Mo-FDA, a deterministic multi-objective extension of the Fractal Decomposition Algorithm using Tchebycheff scalarization and container-based parallelism, ranks best among six algorithms on three of four standard…","keywords":["multi-objective optimization","fractal decomposition algorithm","Tchebycheff scalarization","deterministic metaheuristic","container-based parallel computing","Pareto front approximation","ZDT benchmarks","DTLZ benchmarks"],"falsifier":"Recompute average ranks on the same eight ZDT/DTLZ functions with a single stated evaluation budget and a stated value of $n$ for all algorithms; the central empirical claim is falsified if Mo-FDA no longer ranks first on Hypervolume, GD, and Spread under those reproducible conditions.","tokens_in":3720,"feed_emoji":"🧩","tokens_out":14562,"duration_ms":132106,"temperature":0.7,"pith_summary":"The paper sets out to establish that a deterministic, non-evolutionary solver can handle multi-objective problems competitively once the vector objective is scalarized. It extends the Fractal Decomposition Algorithm with weighted Tchebycheff combinations, solving one scalar problem per weight vector and assembling the results into a Pareto front. To keep the cost of many scalar runs manageable, the paper runs the independent instances as containers that can spread across hosts without algorithmic change. On eight ZDT/DTLZ functions, Mo-FDA places first among six algorithms in average rank-sum scores on Hypervolume, Generational Distance, and Spread, and the container architecture lowers average runtime from 0.8 seconds on one host to 0.5 seconds on two. A careful reader would care because decomposition-based multi-objective methods are often limited by computation cost, and this architecture directly targets that bottleneck.","feed_headline":"Fractal solver beats 5 rivals on 3 of 4 benchmark metrics","feed_subtitle":"Container-parallel solver cuts the same runs from 0.8 to 0.5 seconds on two hosts.","key_machinery":"The load-bearing mechanism is scalarization by the weighted Tchebycheff function $$\\min_{x} \\max_{i=1,\\ldots,k} \\omega_i (f_i(x)-z_i^*)$$ which converts a multi-objective problem into one scalar problem per weight vector $\\omega$, with $z_i^*$ the individual optimum of the $i$-th objective. The geometric carrier is the fractal decomposition: a hypersphere is split into $2D$ sub-hyperspheres at each level, the most promising child is retained, and the pattern repeats until fractal depth $k=5$, after which an intensification local search evaluates three points per dimension from the center of the current hypersphere. The architectural carrier is the container: each scalarized instance is an independent task, so $n$ tasks can run side by side on one host or across hosts without code change, and that independence is what makes the reported time reduction possible.","core_discovery":"On its own terms, the central discovery is that a deterministic search built from hypersphere decomposition and local refinement can stand beside evolutionary multi-objective algorithms once the objective vector is scalarized by the Tchebycheff function. Mo-FDA divides the search space into hyperspheres rather than hypercubes, recursively decomposes the most promising region to depth $k=5$, and refines it with an intensification local search along each dimension. Each of $n$ weight combinations yields one solution, and assembling those runs gives the Pareto front. In the reported comparison, Mo-FDA receives average rank-sum scores of $1.875$ for Hypervolume, $2.25$ for GD, $2.125$ for IGD, and $1.75$ for Spread, placing it first on three metrics and tied for first on Spread, while MEOA/D ranks first on IGD. The parallel layer is a separate claim: the same implementation runs in containers, and two hosts reduce the average time from 0.8 s to 0.5 s compared with one host.","pith_inferences":["I infer that the reported time comparison measures deployment throughput, not algorithmic speedup, because the paper does not state the number of weight combinations $n$ or the per-instance evaluation budget.","A testable extension not pursued in the paper is to hold the weight distribution fixed and compare Mo-FDA against weight-vector decomposition baselines, which would isolate whether the hypersphere geometry or the scalarization choice drives the quality advantage.","The determinism of Mo-FDA could make it a convenient reproducible reference for auditing stochastic multi-objective optimizers, though the paper does not make this application.","If $n$ equals the number of containers, the resource demand grows with the requested front size, so the scalability claim would be sharper if the paper reported the container-to-host ratio used in the 0.5 s result."],"forward_implications":["If the benchmark ranks hold, deterministic decomposition-based solvers are credible alternatives to evolutionary algorithms on standard two- and three-objective problems, at least on the tested metrics.","Because each weight vector generates an independent scalar problem, the container strategy implies that runtime scales down as hosts are added, up to the number of requested Pareto-front points.","The IGD result is a direct corollary: Mo-FDA can lag on inverted generational distance, so users wanting both convergence and diversity should treat IGD as its weaker quality metric.","The independence of instances means the same implementation can move from many containers on one machine to several machines with no algorithmic change, which makes the method adaptable to clusters and cloud environments."],"supporting_citations":[{"why":"Supplies the base Fractal Decomposition Algorithm, including the hypersphere divide-and-conquer and the intensification local search that Mo-FDA adapts.","marker":"[5]"},{"why":"Supplies the weighted Tchebycheff scalarization formulation used to reduce each multi-objective problem to one scalar run per weight vector.","marker":"[4]"},{"why":"Supplies the principal experimental settings for the comparison, which the paper reuses without restating.","marker":"[3]"},{"why":"Provides the comparison harness and the baseline algorithm implementations used to produce the benchmark results.","marker":"[6]"},{"why":"Provides the ZDT benchmark functions used to test Mo-FDA.","marker":"[2]"},{"why":"Provides the DTLZ benchmark functions and the earlier comparison setup that defines the test problems.","marker":"[9]"},{"why":"Supplies the constrained-decomposition-with-grids baseline algorithm that Mo-FDA is compared against.","marker":"[1]"},{"why":"Supplies the global WASF-GA baseline algorithm that Mo-FDA is compared against.","marker":"[8]"}],"fun_headline_variants":["Fractal decomposition tops 3 of 4 multi-objective metrics","Mo-FDA ranks first on 3 benchmarks, tied on Spread","Container-parallel fractal solver cuts runtime 37% on two hosts","Hypersphere MO solver tops 3 metrics, ties Spread"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Mo-FDA and the five comparison algorithms were run under equivalent, fair conditions: the paper borrows experimental settings from another study and never states the maximum number of evaluations or the number of weight combinations $n$, so if those unstated settings favored Mo-FDA, the benchmark advantage would not be established.","fun_headline_variants_meta":{"raw":{"variants":["Fractal decomposition tops 3 of 4 multi-objective metrics","Mo-FDA ranks first on 3 benchmarks, tied on Spread","Container-parallel fractal solver cuts runtime 37% on two hosts","Hypersphere MO solver tops 3 metrics, ties Spread"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001311,"raw_usage":{"total_tokens":5302,"prompt_tokens":864,"completion_tokens":4438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":4371}},"tokens_in":480,"tokens_out":4438,"duration_ms":39989,"temperature":1.0,"reasoning_tokens":4371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:18:14.040386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute average ranks on the same eight ZDT/DTLZ functions with a single stated evaluation budget and a stated value of $n$ for all algorithms; the central empirical claim is falsified if Mo-FDA no longer ranks first on Hypervolume, GD, and Spread under those reproducible conditions.","supporting_citations":[{"cited_title":"Deterministic metaheuristic based on fractal decomposition for large-scale optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the base Fractal Decomposition Algorithm, including the hypersphere divide-and-conquer and the intensification local search that Mo-FDA adapts."},{"cited_title":"Miettinen, F","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted Tchebycheff scalarization formulation used to reduce each multi-objective problem to one scalar run per weight vector."},{"cited_title":"Jiang, J","cited_arxiv_id":null,"evidence_quote":"Supplies the principal experimental settings for the comparison, which the paper reuses without restating."},{"cited_title":"Nebro, J","cited_arxiv_id":null,"evidence_quote":"Provides the comparison harness and the baseline algorithm implementations used to produce the benchmark results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ZDT benchmark functions used to test Mo-FDA."},{"cited_title":"Zitzler, K","cited_arxiv_id":null,"evidence_quote":"Provides the DTLZ benchmark functions and the earlier comparison setup that defines the test problems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the constrained-decomposition-with-grids baseline algorithm that Mo-FDA is compared against."},{"cited_title":"Saborido, A.B","cited_arxiv_id":null,"evidence_quote":"Supplies the global WASF-GA baseline algorithm that Mo-FDA is compared against."}],"review_version":1}