Theoretical Advantage of Multiobjective Evolutionary Algorithms for Problems with Different Degrees of Conflict

Abstract: The field of multiobjective evolutionary algorithms (MOEAs) often emphasizes its popularity for optimization problems with conflicting objectives. However, it is still theoretically unknown how MOEAs perform for different degrees of conflict, even for no conflicts, compared with typical approaches outside this field. As the first step to tackle this question, we propose the OneMaxMin$_k$ benchmark class with the degree of the conflict $k\in[0..n]$, a generalized variant of COCZ and OneMinMax. Two typical non-MOEA approaches, scalarization (weighted-sum approach) and $ε$-constraint approach, are considered. We prove that for any set of weights, the set of optima found by scalarization approach cannot cover the full Pareto front. Although the set of the optima of constrained problems constructed via $ε$-constraint approach can cover the full Pareto front, the general used ways (via exterior or nonparameter penalty functions) to solve such constrained problems encountered difficulties. The nonparameter penalty function way cannot construct the set of optima whose function values are the Pareto front, and the exterior way helps (with expected runtime of $O(n\ln n)$ for the randomized local search algorithm for reaching any Pareto front point) but with careful settings of $ε$ and $r$ ($r>1/(ε+1-\lceil ε\rceil)$). In constrast, the generally analyzed MOEAs can efficiently solve OneMaxMin$_k$ without above careful designs. We prove that (G)SEMO, MOEA/D, NSGA-II, and SMS-EMOA can cover the full Pareto front in $O(\max{k,1}n\ln n)$ expected number of function evaluations, which is the same asymptotic runtime as the exterior way in $ε$-constraint approach with careful settings. As a side result, our results also give the performance analysis of solving a constrained problem via multiobjective way.

• The paper demonstrates that MOEAs achieve full Pareto front coverage in the ONEMAXMINK benchmark without the complex parameter tuning required by classical methods.
• It presents a rigorous runtime analysis showing that MOEAs match the efficiency of ε-constraint approaches while offering simpler implementation and consistent performance.
• The findings imply that for multiobjective pseudo-Boolean optimization under varying degrees of conflict, MOEAs provide a more robust and practical alternative to scalarization and penalty-based methods.

Summary of "Theoretical Advantage of Multiobjective Evolutionary Algorithms for Problems with Different Degrees of Conflict" (2408.04207)

Introduction and Motivation

The paper investigates the theoretical efficiency and coverage properties of Multiobjective Evolutionary Algorithms (MOEAs) in relation to classical non-MOEA approaches—specifically, scalarization (weighted-sum) and $\epsilon$-constraint methods—under varying degrees of objective conflicts in bi-objective pseudo-Boolean optimization. The work addresses two central theoretical gaps: (1) how MOEAs compare against non-MOEA constructs in achieving coverage on the Pareto front as the degree of objective conflict varies—including the trivial “no conflict” case—and (2) the underlying algorithmic difficulties and parameter sensitivities inherent in non-MOEA approaches for such problems.

To systematically analyze these aspects, the ONEMAXMINK benchmark class is proposed, generalizing common bi-objective problems (ONEMINMAX, COCZ) by parameterizing over the conflict degree $k \in [0, n]$. ONEMAXMINK interpolates from full objective coincidence ($k = 0$) to complete conflict ($k = n$), thus affording a granular lens on performance variation across the conflict spectrum.

Benchmark Design and Analytical Results

The ONEMAXMINK problem is defined such that its Pareto front contains $k + 1$ points, directly reflecting the conflict degree. Fundamental theoretical results demonstrate that:

• For any set of scalarization weights, the corresponding set of single-objective optima cannot achieve full coverage of the ONEMAXMINK Pareto front when $k > 2$. Even with arbitrary weight choices, at most three unique Pareto optimal points are found. This limitation extends to both the discrete Boolean and the continuous relaxation of the domain, exposing a fundamental incapacity of the scalarization approach even in convex Pareto setups.
• The $\epsilon$-constraint method can, in principle, generate optima for each Pareto front point, but practical realization demands a careful, instance-dependent selection of $\epsilon$ thresholds and penalty coefficients. Specifically, when employing an exterior penalty function approach, for each Pareto front point, a uniquely tuned parameter is necessary to guarantee optimal solutions, and the required penalty coefficient must meet technical lower bounds. The nonparameter penalty function variant is provably unable to guarantee complete Pareto coverage.
• For $\epsilon$-constraint realization through penalty methods, randomized local search (RLS) achieves coverage in $O(\max\{k,1\}n\log n)$ expected function evaluations, contingent on meticulous parameter selection.

Theoretical Runtime Analysis of MOEAs

Comprehensive runtime bounds are established for (G)SEMO, MOEA/D, NSGA-II, and SMS-EMOA:

• These MOEAs, with parameters set according to standard theoretical guidelines (e.g., population size a small function of $k$), do not require problem-specific parameter tuning or decomposition targeting particular Pareto points.
• All MOEAs analyzed cover the full Pareto front in $O(\max\{k,1\} n \log n)$ expected function evaluations, matching the best-case runtime of the exterior $\epsilon$-constraint approach but with considerably less algorithmic and implementation complexity. For $k = 0$ (no conflict), the runtime matches the corresponding single-objective problem.
• Once a Pareto front point is reached by these MOEAs, their population management dynamics guarantee that such points are preserved in subsequent generations, maintaining diversity and preventing loss of coverage.
• NSGA-II and SMS-EMOA require only moderate population sizes (proportional to the Pareto front size) for these guarantees, linking practical parameterization to rigorous coverage results.

Implications and Discussion

This work formally establishes that, within the ONEMAXMINK framework, MOEAs exhibit robust, parameter-insensitive coverage of the complete Pareto front, even as the degree of conflict varies. In contrast, non-MOEA methods—scalarization and $\epsilon$-constraint—are shown to be fundamentally limited (in the former case) or dependent on highly problem-specific and delicate parameter tuning (in the latter) for full Pareto coverage.

A secondary but notable implication is that representing a standard constrained single-objective problem as a multiobjective instance—with the constraint as an additional objective—renders MOEA-based solution more straightforward to realize, although typically at a runtime overhead proportional to the maximal number of incomparable (Pareto optimal) solutions.

This result suggests two directions for future research and application:

1. Algorithm Selection Guidance: Practitioners addressing multiobjective problems with unknown or variable degrees of conflict should favor MOEAs over scalarization or classical constraint-handling approaches when Pareto front coverage is important, especially in discrete search spaces.
2. Theory of Constrained Optimization via Multiobjectivization: There is further scope to investigate the generalization of these findings, particularly the runtime vs. simplicity tradeoffs in solving (complicated) constrained problems by reformulation into multiobjective terms.

Conclusion

The work rigorously shows, through parameterized analytical results, that MOEAs possess inherent advantages in Pareto front identification for multiobjective pseudo-Boolean optimization across all degrees of objective conflict, without recourse to the intricate parameter tuning demanded by classical non-MOEA approaches. These findings have significant import in both theoretical evolutionary computation and the practical algorithm engineering of multiobjective optimizers.