- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes SURFACE, a method for placing a finite set of Pareto solutions so that their distribution along the Pareto front matches a user-specified preference distribution. Starting from Tchebycheff (or weighted-sum) scalarization, SURFACE reconstructs the weight-to-front mapping from the current scalarized solutions. From this mapping it estimates how much front area each region of the weight simplex represents. It then redistributes the weights through mass equidistribution, using inverse-CDF transport for two objectives and Moser transport for more. The authors give checkable conditions under which Tchebycheff scalarization yields a smooth, non-folding front parameterization. They also prove geometric convergence to a finite-resolution error floor under warm-started inexact solves.
Critical strengths
- The formulation is clean and practically meaningful. Treating the desired front allocation as an explicit input, rather than an implicit consequence of hypervolume, crowding, or reference vectors, gives a principled target. It connects a priori preference specification with a posteriori front representation.
- The method is a natural generalization of SURF to more than two objectives and to nonuniform preferences. It never differentiates through the inner optimizer, keeps all scalarized solves parallel and warm-startable, and works with off-the-shelf solvers.
- The theoretical analysis is careful and complete for its regime.
Critical weaknesses
- Unfair benchmark comparison. SURFACE and Uniform report close numbers on ZDT1/ZDT4 and on DTLZ2/3/4. This suggests the scalarized solver finds near-exact Pareto points even on multimodal problems, while the evolutionary baselines run end-to-end and degrade on exactly those problems. The gains may therefore reflect solver strength rather than weight redistribution. Please report the inner solver, initialization, and compute budget for all methods. Against Uniform, the fair comparison, the improvements are modest.
- Reporting inconsistencies. Section 5.2 refers to 11 settings, but Figure 6 ranks only nine. The excluded four-objective tasks are where SURFACE loses on its own target metric to LMPFE and DEA-GNG. On Resource Gathering, SURFACE’s hypervolume is ten times lower than the baselines’, yet its IGD is far better. Is IGD measured against the KL-regularized front rather than the true one?
Questions for authors
Please see weaknesses.
2 comments
Follow-up: Summary of Responses and Revisions
We thank the reviewer again for their careful evaluation. We hope our corrected benchmark rankings, clarification of the shared inner-solver comparisons, and additional objective-aligned, budget-matched experiments address the reviewer's principal concerns regarding experimental fairness and reporting consistency. As the discussion period concludes, we would sincerely appreciate the reviewer revisiting these results and reconsidering their assessment.
Clarification: Benchmark Fairness, Computational Budgets, and Evaluation Consistency
We thank the reviewer for recognizing
SURFACE's practically meaningful preference formulation and careful, complete theoretical analysis. We provide detailed answers to the concerns as below.W1. Benchmark Fairness and Inner-Solver Budgets
We provide detailed computational accounting and additional wall-clock comparisons in General Response G.1 at the top of this discussion page. Importantly, SURFACE and Uniform share the same exact ZDT/DTLZ oracle, while iterative solvers permit matched update budgets, isolating adaptive weight redistribution. SURFACE performs K inner steps per weight over T rounds, while Uniform allocates KT steps to each fixed weight. Both return the same number of solutions. Other baselines use different end-to-end optimization procedures, including population-based search and auxiliary neural-network training in UMOD. As detailed in G.1, the original Resource Gathering comparisons allow several evolutionary baselines substantially longer execution times. Our additional wall-clock-matched comparison further shows that SURFACE achieves the best preference matching, IGD, and HV among the compared methods. For InnerSolver and initialization, Appendix E details the experimental procedures. The ZDT/DTLZ benchmarks use the same Tchebycheff-scalarized optimization procedure for SURFACE and Uniform. The controlled inexact-solve experiment uses warm-started gradient descent, Deep Sea Treasure, Resource Gathering, and Fruit Tree use KL-regularized soft-value iteration, and Minecart uses a common frozen GPI-LS policy set. SURFACE initializes weights on a regular reference mesh and warm-starts each scalarized problem from the preceding round. Stochastic baselines use five seeds with matched returned-solution cardinality. We will clarify the numerical solver configurations, initialization, and computational budgets for each method in Appendix E.
As in General Response G.2, SURFACE's contribution is adaptive weight redistribution, not a stronger inner optimizer. Its modular framework supports both exact and inexact inner optimization, while PF interpolation and MEQ operate in the (M−1)-dimensional weight space without differentiating through the inner solver. This makes SURFACE compatible with existing application-specific optimizers. IGD is a complementary metric measuring PF approximation quality, rather than SURFACE's directly targeted preference-mass distribution. Thus, modest IGD improvements over Uniform are expected given the same inner solver and uniform target preference. SURFACE directly targets preference-mass distribution across the PF. For example, it reduces the PF-area residual from 0.505 to 0.316 on ZDT4 and from 1.078 to 0.693 on M=3 DTLZ3 using the same inner solver. The similar results on DTLZ2/3/4 are consistent with their shared PF geometry. These improvements arise from weight redistribution under the same scalarized optimization procedure.
W2. Reporting Consistency and HV–IGD Interpretation
We thank the reviewer for identifying the inconsistency between the 11 settings described in Section 5.2 and the nine shown in the original Figure 6, which we have corrected by including the four-objective cases where LMPFE and DEA-GNG outperform SURFACE on individual metrics. Across all 11 settings, SURFACE retains the best mean PF-area imbalance and IGD ranks (both 1.18), compared with 3.18 and 2.27 for the second best. This correction concerns experimental reporting and does not affect the underlying results, theoretical guarantees, or conclusions.
For Resource Gathering, we clarify that both IGD and HV are evaluated using unregularized environment returns, not the KL-regularized training objectives (Appendix E.4). These metrics characterize different aspects of PF representation: IGD evaluates proximity and coverage relative to the PF reference, while HV measures dominated objective-space volume. The original evolutionary baselines were allowed longer runtimes and optimized their native objectives, whereas SURFACE used KL-regularized policy optimization. These differences can contribute to the higher evolutionary HV in the original comparison. As detailed in General Response G.1, we additionally aligned the KL-regularized objectives across methods and evaluated their unregularized returns under matched computational budgets. Under the same time budget, SURFACE achieves the best HV, IGD, and preference residual. Moreover, SURFACE's HV remains higher than that of the evolutionary baselines even when they are allowed to complete their longer runs under the aligned objective setting. We will incorporate and clarify the these in Appendix E.4.
Together, these clarifications and additional experiments demonstrate the effectiveness of preference-aware weight redistribution independently of inner-solver strength, with consistent improvements under controlled and computationally aligned comparisons.