Abstract
Constraint handling is one of the most influential aspects of applying metaheuristics to real-world applications, which can hamper the search progress if treated improperly. In this work, we focus on a particular case—the box constraints—for which many boundary constraint handling methods (BCHMs) have been proposed. We call for the necessity of studying the impact of BCHMs on metaheuristics' performance and behavior, which receives seemingly little attention in the field. We target quantifying such impacts through systematic benchmarking by investigating 28 major variants of Differential Evolution (DE) taken from the modular DE framework and 13 commonly applied BCHMs, resulting in 364 algorithm instances. After executing these instances on the BBOB/COCO problem set, we analyze the best-reached objective function value and the percentage of repaired solutions (PORS). Our results clearly show that the choice of BCHMs substantially affects empirical performance and the number of generated infeasible solutions, allowing us to provide general guidelines for selecting an appropriate BCHM.
Introduction: The "Invisible" Boundary Problem
Differential Evolution (DE) is a robust tool for continuous optimization, yet its core mechanism—the mutation operator—frequently generates "trial vectors" that fall outside the feasible search space. While researchers dedicate significant effort to tuning mutation and crossover strategies, the method used to "repair" these out-of-bounds solutions is often treated as a minor implementation detail.
This study argues that Boundary Constraint Handling Methods (BCHMs) are not just secondary settings; they introduce structural bias and fundamentally alter search trajectories. By systematically quantifying these impacts, we demonstrate that the choice of repair strategy can be as critical as the choice of the optimizer itself.
Methodology
To evaluate the interaction between DE variants and repair strategies, we utilized a modular DE framework to construct 364 unique algorithm instances. These instances were formed by combining:
- 14 Mutation Schemes: Including
rand/1,best/2,NSDE, andTrigonometricvariants. - 2 Crossover Operators: Binomial and Exponential.
- 13 BCHMs: Ranging from standard methods like Projection (clamping) and Reflection to more specialized strategies like Conservatism and Resampling.
Each instance was benchmarked using the BBOB/COCO suite in 30 dimensions, covering 24 functions across five distinct problem groups (Separable, Low/Moderate Conditioning, High Conditioning/Unimodal, and two Multi-modal groups).
Behavioral Impact: The Percentage of Repaired Solutions (PORS)
A key metric in this study is the PORS, which measures how often an algorithm generates infeasible solutions that require repair. Our analysis reveals that behavioral impact is highly sensitive to the crossover operator. DE variants using exponential crossover consistently generated more infeasible solutions than those using binomial crossover, making them significantly more dependent on the choice of BCHM.
In extreme cases, such as the "Linear Slope" function, the PORS exceeded 90%. This indicates that the algorithm spends the majority of its computational budget searching along the boundaries, effectively transforming the search into a boundary-constrained optimization problem.
Performance Analysis and Robustness
The choice of BCHM can move an algorithm from the top of the leaderboard to the bottom within the same function group. We observed that high PORS is not inherently detrimental; it often indicates high exploration near search limits. However, the repair method must facilitate effective search progress rather than trapping the population in local optima near the boundary.
Key Findings:
- Top Performers: Resampling and Conservatism emerged as the most robust BCHMs. Resampling (trying a new mutation until a feasible point is found) was particularly effective for binomial crossover, while Conservatism (retaining the base vector if the mutant is infeasible) excelled in many exponential crossover configurations.
- Crossover Dominance: The crossover operator is the primary determinant of which BCHM will perform optimally.
- Reliable Fallbacks: For general applications, Resampling is recommended for binomial crossover, while Projection Midpoint serves as a reliable second option for exponential crossover.
Conclusion and Future Work
This research highlights that BCHMs are a critical component of the DE pipeline. The significant performance swings observed across different repair strategies suggest that these methods should be explicitly reported and carefully selected in optimization studies.
Future work will explore the impact of BCHMs in higher dimensions, where the search space volume is increasingly concentrated near boundaries, and investigate "Adaptive BCHMs" that can dynamically switch repair strategies during a run to optimize search behavior.
Summary updated