MOEA/D with Uniformly Randomly Adaptive Weights

Aluizio F. R. Ara\'ujo; Hansenclever F. Bassani; Lucas R. C. de Farias; Pedro H. M. Braga

REVIEW 3 major objections 6 minor 12 references

MOEA/D with Uniformly Randomly Adaptive Weights

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read MOEA/D-URAW claims that adapting subproblem weights to the sparsity of the population lets a decomposition-based optimizer match or beat fixed-weight variants in 77.5% of tested problems while keeping population size flexible.

desk verdict Useful incremental algorithm paper with a real empirical flaw at 6 objectives: the DD baseline is invalid, and the headline 77.5% rests partly on that comparison. read the letter →

arxiv 1908.05383 v1 pith:UVHNLNML submitted 2019-08-15 cs.NE

Lucas R. C. de Farias , Pedro H. M. Braga , Hansenclever F. Bassani , Aluizio F. R. Araújo This is my paper · ORCID

classification cs.NE

keywords Many-objectiveoptimizationMulti-objectiveEvolutionstrategiesDecompositionmethodsAdaptiveweightsSparsitylevelMOEA/DHypervolume

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Decomposition-based multi-objective optimizers split a problem into scalar subproblems, one per weight vector, and the final solution set depends heavily on how those weights are chosen. This paper proposes MOEA/D-URAW, which starts from a uniformly random set of weights and then, during the run, removes subproblems whose solutions are crowded and adds subproblems where the population is sparse in objective space. The sparsity measure is the product of the distances to the nearest neighbors, one distance per objective. The intended payoff is twofold: population size is no longer tied to the combinatorial growth of fixed simplex grids, and the weight distribution can follow the actual Pareto front instead of assuming a shape in advance. Across the WFG41–48 problems with two to six objectives, the paper reports statistically better hypervolume than three fixed-weight MOEA/D variants in 72.5% of cases, equal in another 5%, and worse in the remaining 22.5%.

What carries the argument

The load-bearing object is the sparsity level $SL(ind_j, pop) = \prod_{i=1}^{m} L^{NN_j^i}_2$, the product of the $m$ nearest-neighbor Euclidean distances of an individual in objective space, where $m$ is the number of objectives. A low value marks a crowded subproblem to delete; a high value computed over the external archive marks a gap where a new subproblem should be created. The new weight vector is generated from the objective values of the sparsest external individual by inverting the Tchebycheff scalarizing condition. Uniformly Randomly initialization supplies the flexible population size, and the WS-transformation maps generated weights into solution directions before the run.

What would settle it

Run the same 100-run hypervolume comparison on a suite containing only strong-concave (WFG43) and strong-convex (WFG44) fronts across two to six objectives; if MOEA/D-URAW is not statistically better than its own fixed-weight uniform-random initialization on either shape, the sparsity criterion is not a general geometry-free guide. The paper's reported WFG43 and WFG44 results already point in that direction.

Watch

Extended reading notes

Core claim

The paper's central claim is that a decomposition-based multi-objective optimizer can improve its final solution set by letting the weight vectors follow the population's sparsity instead of being fixed before the run. Starting from a Uniformly Randomly generated set of subproblems, the algorithm repeatedly removes the subproblem whose solution is most crowded and inserts a new subproblem at the least crowded point of the external archive, with the new weight vector computed from that point's objective values. On the WFG41–48 benchmark class, 2–6 objectives, 100 runs each, this scheme gives hypervolume results that are statistically better than all three fixed-weight baselines in 72.5% of cases and tied in 5%, for 77.5% better-or-equal overall. The paper explicitly reports that the mechanism fails to improve on strong-concave and strong-convex fronts (WFG43 and WFG44) and leaves one disconnected-concave case (WFG47, 2 objectives) where a fixed-weight method wins.

Load-bearing premise

The method assumes that the product of nearest-neighbor distances in objective space reliably shows where the population is crowded or sparse on every Pareto front shape, and the paper's own strongly concave and strongly convex results show this assumption is not universally true.

Editorial extensions

If this is right

MOEA/D-URAW decouples population size from the combinatorial growth of fixed simplex weight sets, so runs with 4, 5, and 6 objectives can use the same style of weight generation.
Because weights are re-drawn from the current population's sparse regions, the algorithm does not need a prior assumption about the Pareto front's geometry for most tested shapes.
Across WFG41–48 with 2–6 objectives, the adaptive version achieves higher average hypervolume than the three fixed-weight variants in 72.5% of cases and ties in 5%, for a 77.5% better-or-equal rate.
On strongly concave (WFG43) and strongly convex (WFG44) fronts, fixed-weight versions win, so the sparsity criterion is not reliable for those geometries.
The frequency of weight updates (every 5% of generations, none in the last 10%) is a working choice, not a settled parameter; the paper identifies no consensus on the best update schedule.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the sparsity criterion is the limiting factor rather than the initial weight generation, then a modified sparsity measure that normalizes distances per objective or accounts for the local curvature of the front could extend the method to the strong-concave and strong-convex cases where it currently loses.
The flexible-population property suggests a natural test the paper does not run: start with a small subproblem set and let the adaptive additions alone grow the population, which would directly measure how much of the gain comes from adaptation rather than from the initial uniform-random spread.
The WFG47 anomaly—fixed Tchebycheff weights winning on a disconnected concave front—could be diagnosed by checking whether the sparsity-driven insertions place subproblems inside the gap between front components, where they can never produce nondominated solutions.
A practical extension would pair URAW's weight adaptation with other scalarizing functions, since the paper's experiments use only Tchebycheff decomposition; the sparsity logic itself is independent of the scalarizing function.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick take: this is a decent, honest incremental paper about adaptive weights in MOEA/D, but the headline number is partly built on an invalid baseline at six objectives. If you read it, treat the 77.5% claim with caution; the algorithm itself is worth a look.

What's new: the URAW combination — UR weight initialization plus sparsity-based add/remove of subproblems, using the external population to spawn new ones, and skipping AWA's re-association step. The components come from prior work (Zhang et al. for UR initialization, Qi et al. for AWA), but I don't know of another paper that puts them together this way, and the flexible population size is a nice side effect. The pseudocode is clear, the Wilcoxon tests and standard deviations are reported consistently, and the authors are explicit that the adaptation doesn't help on strong concave/convex PFs (WFG43/44). That honesty deserves credit.

The soft spots, in order of severity. First, the stress-test checks out: for m=6 objectives, N=126 gives H=4, which is less than m. So the 'Das and Dennis' weight set used in Tables 5–6 is entirely boundary weights with zero components, and the WS-transformation in Eq. (4) divides by λ_i. The paper never states how this is handled, so the 6-objective DD comparator is not the method described in Sec. 2.2.1. Since six-objective runs are 8 of the 40 comparisons in the 77.5% tally, that number is not grounded as stated. The comparison against UR (no adaptation) and TSF is still meaningful at 6 objectives, but the DD column there is suspect.

Second, no comparison with MOEA/D-AWA, the direct adaptive predecessor. The paper says the adaptation method is 'based on' AWA, so omitting it from the experiments is a real gap. Third, the adaptation frequency (every 5% of generations, no changes in the last 10%) was chosen from 'preliminary evaluation' on the same benchmarks; that's tuning on the test set. It's not fatal, but it makes the 'state-of-the-art' generalization weaker. No code is provided.

Who's this for: people working on decomposition-based EMO who want a simple adaptive-weight recipe with a flexible population size. It's a workshop-grade incremental contribution, not a breakthrough. My own verdict would be: accept after major revision, with the six-objective DD baseline fixed or removed, AWA added to the comparison, and ideally code released. A serious referee should engage with it — there is a real idea here — but the current numeric claim needs repair.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes MOEA/D-URAW, a decomposition-based many-objective optimizer that initializes weight vectors with the Uniformly Random (UR) method and then periodically adds and removes subproblems according to a sparsity measure computed from nearest-neighbor distances in objective space. It is evaluated on eight WFG test problems (WFG41–48) with 2–6 objectives against three fixed-weight instantiations of MOEA/D: Das–Dennis (DD), Uniformly Random (UR), and Tchebycheff Scalarizing Function (TSF). Hypervolume results over 100 runs, tested with the Wilcoxon rank-sum test at the 5% level, are claimed to show statistically better performance in 72.5% of cases and better-or-equal performance in 77.5%.

Significance. If sustained, the paper offers a simple and flexible alternative to fixed uniform weights: the population size can be an arbitrary integer and the weights can adapt to Pareto-front geometry without explicitly estimating the front shape. The experimental protocol is generally reasonable: means and standard deviations over 100 runs are reported, a non-parametric test is used, and the authors honestly report the WFG43 and WFG44 cases where adaptation hurts. The main weaknesses are that the 6-objective DD baseline is internally invalid, the most relevant adaptive baseline (MOEA/D-AWA) is missing, and several key parameters of the adaptation mechanism are chosen without a reported sensitivity analysis. These issues directly affect the credibility of the headline 77.5% claim.

major comments (3)

[Section 2.2.1, Table 1, Eq. (4)] For m=6, the choice N=126 forces H=4 (since C(4+6-1,6-1)=C(9,5)=126), violating the requirement stated in Section 2.2.1 that H must be no smaller than m to prevent intermediate points being created. All DD weight vectors then contain at least one zero component, and the WS-transformation in Eq. (4) is undefined (division by zero) for every such vector; the same applies to the axis vectors included in the UR initialization. Consequently, the 'DD' results for 6 objectives reported in Tables 5 and 6 are not for the method described in Section 2.2.1 and may represent a handicapped baseline. Since the 6-objective cases contribute 8 of the 40 comparisons entering the 77.5% aggregate, the central quantitative claim is not currently grounded in the many-objective regime that the paper explicitly targets. The authors should either use a valid Das–Dennis set (e.g., H=5, N=252, with an equivalent evaluation budget) or exclude these comparisons and recompute the aggregate statistics.
[Sections 3.2 and 4.2] The adaptive mechanism of MOEA/D-URAW is explicitly based on MOEA/D-AWA (Qi et al., 2014), yet MOEA/D-AWA is not included as a baseline in the experiments. The paper's claim of outperforming 'state-of-the-art methods' is therefore under-supported: the most natural comparator, which also uses sparsity-based addition and removal of subproblems, is absent. The authors should add MOEA/D-AWA (or a faithful reimplementation) to the comparison, or rephrase the conclusions to refer only to the three fixed-weight variants actually tested.
[Section 3.3 and Table 2] The adaptation frequency (every 5% of generations), the blackout period (the last 10% of generations), the number of subproblems added/removed nus, and the external-population size limit (2|P|) are fixed values whose justification is only a 'preliminary evaluation' reported in one sentence. These are free parameters of the proposed method, and no sensitivity analysis is given. Since the paper's positive conclusion depends on these settings, the authors should describe how these values were chosen and provide at least a robustness study over the adaptation frequency and nus.

minor comments (6)

[Abstract and Section 5] The decimal separator is inconsistent: the abstract and Section 5 use '77,5%' while the rest of the paper uses '77.5%'; please unify.
[Algorithm 1, line 26] The condition 'if Gen = Genmax× 5% and < Genmax× 90%' is not a well-formed logical/arithmetic expression. It should be clarified, for example as 'if (Gen mod (0.05·Genmax) == 0) and (Gen < 0.9·Genmax)'. As written, the pseudocode is ambiguous and hurts reproducibility.
[Section 3.1] The text states that 'the population size is flexible, that is, it independs on the number of objectives'; the word 'independs' should be replaced by 'is independent'.
[Table 1] For N=126, nus is defined as 0.05N, which equals 6.3, but the table lists nus=6. Please clarify whether nus is rounded down or otherwise truncated.
[Equation (5)] The notation LN N j i is not fully defined; the indices i and j should be explained so that the product over 'the m closest Euclidean distances' is unambiguous.
[Figure captions] The captions of Figures 1 and 2 refer to 'the best' and 'median' HV values, but it should be stated explicitly whether 'best' means the maximum HV across 100 runs or the run associated with the best hypervolume.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: adaptive weights are a benchmark-tested heuristic, not an input fitted into a prediction; flagged non-circular validity issue in 6-objective DD baseline.

full rationale

MOEA/D-URAW is an empirical algorithm study. The sparsity level (Eq. 5) and the new-subproblem weight formula (Eq. 6) are heuristic rules inherited from MOEA/D-AWA [7]; they are not fitted to the outcomes and then reported as predictions. Performance is assessed against external fixed-weight baselines (DD, UR, TSF) by hypervolume on WFG41-48, so the central comparison is self-contained. The authors honestly flag where the mechanism fails (strong concave/strong convex PFs, WFG43/44) and note that adaptation frequency has no consensus, which weakens the claim but is not circularity. No load-bearing self-citation appears: [6]-[9] are prior independent works by other groups. The only circularity-adjacent item is that the 5% adaptation cadence and 10% freeze were chosen in a preliminary evaluation on the same benchmark suite; this is benchmark tuning and a possible selection-bias risk, but not a definitional reduction. Separately, as a non-circular internal-validity flag: Section 2.2.1 requires H >= m for Das-Dennis weights, while Table 1 sets N=126 for 5 and 6 objectives; for m=6 this implies H=4<m, so the DD weight vectors have zero components and Eq. (4)'s WS-transformation divides by lambda_i=0, making the 6-objective DD baseline not the method described. This concerns the 77.5% tally but does not make the proposed method's derivation circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the sparsity-based adaptation heuristic, the WFG benchmark suite as a stand-in for real Pareto fronts, the hypervolume indicator as the performance metric, and the standard MOEA/D framework. The method introduces no new theoretical entities; its free parameters (adaptation frequency, nus, EP limit) are hand-chosen and one is tuned on the test suite.

free parameters (4)

adaptation_frequency = 5% of generations
Chosen from a preliminary evaluation, per Section 3.2, not systematically varied; affects results.
adaptation_blackout_ratio = 10% (no adaptation in final 10% of generations)
Hand-chosen, no sensitivity analysis.
nus (number of subproblems added/removed) = 6 (0.05N)
Set to 5% of population size, no justification or sensitivity study.
EP_size_limit = 2|P|
External population size cap selected without analysis; impacts which solutions spawn new subproblems.

assumptions (5)

domain assumption The sparsity level formula (product of m nearest-neighbor distances) is an appropriate measure of solution crowding for weight adaptation
Taken from MOEA/D-AWA (ref [7]); the paper's results show it is PF-shape dependent.
domain assumption The WFG41-48 test problems represent the range of relevant Pareto front geometries
The benchmark suite is standard in the MOEA literature, but the reported failures on strong concave/convex shapes show its limits.
domain assumption The hypervolume metric with min-max normalization and reference point (1.2,...,1.2) is a valid comparison measure for different PF shapes
HV is standard, but normalization across all experiments may mask scale differences.
standard math The MOEA/D framework of Li and Zhang (2009) is used as the base algorithm
Standard background; no derivation in this paper.
standard math Tchebycheff decomposition with reference point z* is used to scalarize subproblems
Standard decomposition from Zhang and Li (2007).

how reviews work

0 comments

Cite this review

Pith. "Pith review of MOEA/D with Uniformly Randomly Adaptive Weights." pith.science (2026). https://pith.science/paper/UVHNLNML

@misc{pith2026190805383,
  author       = {Pith},
  title        = {Pith review of: MOEA/D with Uniformly Randomly Adaptive Weights},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVHNLNML}},
  note         = {Machine review of arXiv:1908.05383}
}

read the original abstract

When working with decomposition-based algorithms, an appropriate set of weights might improve quality of the final solution. A set of uniformly distributed weights usually leads to well-distributed solutions on a Pareto front. However, there are two main difficulties with this approach. Firstly, it may fail depending on the problem geometry. Secondly, the population size becomes not flexible as the number of objectives increases. In this paper, we propose the MOEA/D with Uniformly Randomly Adaptive Weights (MOEA/DURAW) which uses the Uniformly Randomly method as an approach to subproblems generation, allowing a flexible population size even when working with many objective problems. During the evolutionary process, MOEA/D-URAW adds and removes subproblems as a function of the sparsity level of the population. Moreover, instead of requiring assumptions about the Pareto front shape, our method adapts its weights to the shape of the problem during the evolutionary process. Experimental results using WFG41-48 problem classes, with different Pareto front shapes, shows that the present method presents better or equal results in 77.5% of the problems evaluated from 2 to 6 objectives when compared with state-of-the-art methods in the literature.

Figures

Figures reproduced from arXiv: 1908.05383 by the authors.

**Figure 2.** Final solutions set with the best (blue circle) and median (green cross sign) HV metric values obtained by the algo [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Indraneel Das and John E Dennis. 1998. Normal-boundary intersection: A new method for generating the Pareto surface in nonlinear multicriteria optimization problems. SIAM Journal on Optimization 8, 3 (1998), 631–657

work page 1998
[2]

Kalyanmoy Deb. 2001. Multi-objective optimization using evolutionary algorithms . Vol. 16. John Wiley & Sons. MOEA/D with Uniformly Randomly Adaptive Weights GECCO ’18, July 15–19, 2018, Kyoto, Japan (a) WFG41 - DD (b) WFG41 - UR (c) WFG41 - TSF (d) WFG41 - URAW (e) WFG42 - DD (f) WFG42 - UR (g) WFG42 - TSF (h) WFG42 - URAW (i) WFG43 - DD (j) WFG43 - UR (...

work page 2001
[3]

Fangqing Gu, Hai-Lin Liu, and Kay Chen Tan. 2012. A multiobjective evolution- ary algorithm using dynamic weight design method. International Journal of innovative Computing, Information and Control 8, 5 (B) (2012), 3677–3688

work page 2012
[4]

Saku Kukkonen and Kalyanmoy Deb. 2006. A fast and effective method for prun- ing of non-dominated solutions in many-objective problems. In PPSN, Vol. 4193. 553–562

work page 2006
[5]

Hui Li and Qingfu Zhang. 2009. Multiobjective optimization problems with complicated Pareto sets, MOEA/D and NSGA-II.IEEE transactions on evolutionary computation 13, 2 (2009), 284–302

work page 2009
[6]

Miqing Li and Xin Yao. 2017. What Weights Work for You? Adapting Weights for Any Pareto Front Shape in Decomposition-based Evolutionary Multi-Objective Optimisation. arXiv preprint arXiv:1709.02679 (2017)

work page Pith review arXiv 2017
[7]

Yutao Qi, Xiaoliang Ma, Fang Liu, Licheng Jiao, Jianyong Sun, and Jianshe Wu

work page
[8]

Rui Wang, Robin C Purshouse, and Peter J Fleming. 2015. Preference-inspired co- evolutionary algorithms using weight vectors. European Journal of Operational Research 243, 2 (2015), 423–441

work page 2015

Show all 12 references

[9]

Mengyuan Wu, Sam Kwong, Yuheng Jia, Ke Li, and Qingfu Zhang. 2017. Adap- tive weights generation for decomposition-based multi-objective optimization using Gaussian process regression. In Proceedings of the Genetic and Evolutionary Computation Conference. ACM, 641–648

2017
[10]

Qingfu Zhang and Hui Li. 2007. MOEA/D: A multiobjective evolutionary algo- rithm based on decomposition. IEEE Transactions on evolutionary computation 11, 6 (2007), 712–731

2007
[11]

Qingfu Zhang, Wudong Liu, and Hui Li. 2009. The performance of a new ver- sion of MOEA/D on CEC09 unconstrained MOP test instances. In Evolutionary Computation, 2009. CEC’09. IEEE Congress on . IEEE, 203–208. GECCO ’18, July 15–19, 2018, Kyoto, Japan Lucas R. C. de Farias, Ped...

2009
[2014]

Evolutionary computation 22, 2 (2014), 231–264

MOEA/D with adaptive weight adjustment. Evolutionary computation 22, 2 (2014), 231–264

2014

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Indraneel Das and John E Dennis. 1998. Normal-boundary intersection: A new method for generating the Pareto surface in nonlinear multicriteria optimization problems. SIAM Journal on Optimization 8, 3 (1998), 631–657

work page 1998

[2] [2]

Kalyanmoy Deb. 2001. Multi-objective optimization using evolutionary algorithms . Vol. 16. John Wiley & Sons. MOEA/D with Uniformly Randomly Adaptive Weights GECCO ’18, July 15–19, 2018, Kyoto, Japan (a) WFG41 - DD (b) WFG41 - UR (c) WFG41 - TSF (d) WFG41 - URAW (e) WFG42 - DD (f) WFG42 - UR (g) WFG42 - TSF (h) WFG42 - URAW (i) WFG43 - DD (j) WFG43 - UR (...

work page 2001

[3] [3]

Fangqing Gu, Hai-Lin Liu, and Kay Chen Tan. 2012. A multiobjective evolution- ary algorithm using dynamic weight design method. International Journal of innovative Computing, Information and Control 8, 5 (B) (2012), 3677–3688

work page 2012

[4] [4]

Saku Kukkonen and Kalyanmoy Deb. 2006. A fast and effective method for prun- ing of non-dominated solutions in many-objective problems. In PPSN, Vol. 4193. 553–562

work page 2006

[5] [5]

Hui Li and Qingfu Zhang. 2009. Multiobjective optimization problems with complicated Pareto sets, MOEA/D and NSGA-II.IEEE transactions on evolutionary computation 13, 2 (2009), 284–302

work page 2009

[6] [6]

Miqing Li and Xin Yao. 2017. What Weights Work for You? Adapting Weights for Any Pareto Front Shape in Decomposition-based Evolutionary Multi-Objective Optimisation. arXiv preprint arXiv:1709.02679 (2017)

work page Pith review arXiv 2017

[7] [7]

Yutao Qi, Xiaoliang Ma, Fang Liu, Licheng Jiao, Jianyong Sun, and Jianshe Wu

work page

[8] [8]

Rui Wang, Robin C Purshouse, and Peter J Fleming. 2015. Preference-inspired co- evolutionary algorithms using weight vectors. European Journal of Operational Research 243, 2 (2015), 423–441

work page 2015