Abstract
This study targets the mixed-integer black-box optimization (MI-BBO) problem where continuous and integer variables should be optimized simultaneously. The CMA-ES, a population-based stochastic search method, shows excellent performance in continuous BBO. However, when applied to MI-BBO with straightforward discretization, the variance corresponding to integer variables often becomes much smaller than the discretization granularity before reaching the optimal solution, leading to stagnation. This is particularly prevalent with binary variables. We propose a simple extension of the CMA-ES based on lower-bounding the marginal probabilities associated with the generation of integer variables. Numerical experiments demonstrate the efficiency and robustness of the proposed method in both single- and multi-objective cases.
Introduction: The Stagnation Problem in MI-BBO
The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is a premier method for continuous black-box optimization. However, real-world applications often involve Mixed-Integer Black-Box Optimization (MI-BBO), requiring simultaneous handling of continuous and discrete variables.
When standard CMA-ES is applied to these problems via simple rounding, it frequently suffers from stagnation. As the algorithm converges, its step-size ($\sigma$) and variance shrink. If they become significantly smaller than the integer granularity (e.g., the distance between 0 and 1), the algorithm stops exploring different discrete values. It becomes trapped on a "plateau" where small changes in continuous parameters do not affect the discretized output. This "bit fixation" is especially problematic for binary variables.
As shown in the figure above, the failure in binary optimization is often driven by the mean vector ($m$) drifting far from the discretization threshold while the variance is small, making mutations to alternative states statistically impossible.
Proposed Method: CMA-ES with Margin
To overcome these limitations, we introduce CMA-ES with Margin. Inspired by Estimation of Distribution Algorithms (EDAs), the "margin" ensures that the probability of sampling any alternative integer value never drops below a threshold $\alpha$. This prevents the distribution from collapsing onto a single point.
Technical Innovations
- Marginal Probability Lower-Bounding: If the probability of sampling a different integer value falls below $\alpha$, the algorithm shifts the mean closer to the discretization threshold and increases coordinate-wise variance to satisfy the margin.
- Affine Transformation Correction: Instead of modifying rounding logic, the authors adjust the sampling distribution itself by modifying the mean vector and a diagonal transformation matrix before generating candidates.
- Automated Hyperparameters: The paper suggests a robust default margin: $\alpha = (N\lambda)^{-1}$, where $N$ is the dimension and $\lambda$ is the population size.
The method preserves the core CMA-ES logic (parameters shown in the table above), ensuring that covariance matrix updates and step-size adaptation remain functional while preventing distribution collapse.
Multi-Objective Extension
The margin logic was successfully integrated into MO-CMA-ES. The authors proved that this correction preserves the "elitist" properties of multi-objective search, ensuring it does not disrupt the ranking of successful individuals.
Experimental Results
The proposed method was evaluated on synthetic MI-BBO benchmarks, including combinations of Sphere/Ellipsoid functions with OneMax/LeadingOnes.
- Robustness: In single-objective tests, CMA-ES with Margin maintained a 100% success rate up to 60D, whereas the baseline CMA-ES-IM success rate dropped to 8% on binary problems.
- Efficiency: On integer problems like SphereInt, the proposed method reached the optimum with significantly fewer evaluations.
- Multi-Objective Performance: On the DSLOTZ benchmark, the margin-corrected version achieved significantly higher Hypervolume than vanilla MO-CMA-ES and NSGA-II, both of which suffered from bit fixation.
Conclusion and Future Work
CMA-ES with Margin provides a simple, principled solution to stagnation in mixed-integer optimization. While highly effective, the current coordinate-wise margin correction may struggle with highly non-separable, ill-conditioned functions. Future research will explore moving beyond coordinate-wise margins to consider joint distributions and validating the method on complex real-world engineering black boxes.
Summary updated