Cascading Multimodal Optimizers (CMA-ES-DS)

• Cascading Multimodal Optimizers (CMA-ES-DS) are evolutionary algorithms that coordinate multiple CMA-ES instances to yield a top-quality leader while enforcing strict pairwise Euclidean distance among solutions.
• They utilize a cascading initialization and dynamic tabu regions to systematically reject candidates that do not meet the minimum separation criterion, ensuring batch diversity.
• Empirical evaluations reveal that CMA-ES-DS achieves 20–50% lower average loss compared to baselines, reliably generating valid diversified solution batches under varying constraints.

Cascading Multimodal Optimizers (CMA-ES-DS) are a class of evolutionary optimization algorithms tailored for scenarios where generating a batch of input-diverse, high-quality solutions is essential. Central to their design is the simultaneous achievement of two goals: obtaining the best possible solution (“leader”) and ensuring that all solutions within the batch maintain a prescribed minimum pairwise Euclidean distance. This methodology is especially relevant in engineering or design applications where alternative solutions serve as a hedge against unknown constraints or post-hoc user preferences. The Cascading CMA-ES-Diversity Search (CMA-ES-DS) algorithm realizes these objectives by coordinating multiple parallel instances of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), dynamically enforcing “tabu” regions to maintain diversity, and systematically outperforming state-of-the-art multimodal and random sampling baselines in both quality and diversity metrics (Santoni et al., 19 Feb 2025).

1. Problem Formulation and Motivating Context

A frequent real-world requirement in black-box optimization is the generation of $k$-sized solution batches $X^* = \{x^1, ..., x^k\} \subseteq S$ such that:

1. The batch includes the best-found solution, $x^1 = \arg\min_{x \in S} f(x)$, for $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$.
2. All pairwise distances satisfy $\|x^i - x^j\| \ge d_{\min}$, $i \ne j$, for a given $d_{\min}>0$.
3. The batch minimizes the average fitness $\frac{1}{k} \sum_{i=1}^k f(x^i)$.

Prior studies, notably by Santoni et al. [arXiv 2024], demonstrated that existing random sampling and state-of-the-art multimodal optimizers consistently fail to balance quality and strict minimum-distance constraints, sometimes performing no better than uniform random sampling. This performance gap motivates the development of algorithms explicitly designed for high-quality, $d_{\min}$-separated batch extraction (Santoni et al., 19 Feb 2025).

2. Algorithmic Structure and Tabu-Cascading Framework

CMA-ES-DS operates by orchestrating $k$ independent CMA-ES instances indexed $X^* = \{x^1, ..., x^k\} \subseteq S$0, each with its own initial mean, covariance matrix, and tabu region center:

• Initialization (Cascading Sampling): The initial mean $X^* = \{x^1, ..., x^k\} \subseteq S$1 for each instance is sampled uniformly in $X^* = \{x^1, ..., x^k\} \subseteq S$2 with the requirement $X^* = \{x^1, ..., x^k\} \subseteq S$3 for all $X^* = \{x^1, ..., x^k\} \subseteq S$4. This ensures initial spatial separation via cascading rejection sampling. Each instance uses step-size $X^* = \{x^1, ..., x^k\} \subseteq S$5 and covariance $X^* = \{x^1, ..., x^k\} \subseteq S$6.
• Dynamic Tabu Regions: At generation $X^* = \{x^1, ..., x^k\} \subseteq S$7, each instance $X^* = \{x^1, ..., x^k\} \subseteq S$8 samples up to $X^* = \{x^1, ..., x^k\} \subseteq S$9 candidates $x^1 = \arg\min_{x \in S} f(x)$0, rejecting candidates falling within any tabu ball $x^1 = \arg\min_{x \in S} f(x)$1 for $x^1 = \arg\min_{x \in S} f(x)$2.
• Updating Tabu Centers: Upon evaluation, each instance updates its tabu center $x^1 = \arg\min_{x \in S} f(x)$3 to the lowest-fitness candidate in its current valid population.
• Convergence and Restarts: If an instance meets its stopping criteria (e.g., TolX, TolFun), its tabu center is fixed. If all $x^1 = \arg\min_{x \in S} f(x)$4 converge before exceeding the evaluation budget $x^1 = \arg\min_{x \in S} f(x)$5, all are restarted using the cascading initialization process.

Synchronous generational updates and systematic enforcement of tabu regions are critical for maintaining both high batch diversity and quality. The method is described by the following iterative process:

Step	Action	Constraint/Update
Parallel instance creation	Sample $x^1 = \arg\min_{x \in S} f(x)$6, ensure minimum distance to $x^1 = \arg\min_{x \in S} f(x)$7 for $x^1 = \arg\min_{x \in S} f(x)$8	$x^1 = \arg\min_{x \in S} f(x)$9 valid if $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$0
Candidate sampling	Draw $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$1	Reject if $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$2
Tabu region update	Set $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$3 to best candidate found	$f: S\subseteq\mathbb{R}^D \to \mathbb{R}$4

3. Mathematical Formulation and Key Hyperparameters

The core mathematical concepts and hyperparameter settings include:

• Tabu Ball Definition: For center $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$5, $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$6.
• CMA-ES Update (abridged): The standard CMA-ES equations govern mean, covariance, and step-size updates:
  • $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$7
  • $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$8
• Population sizing: $f: S\subseteq\mathbb{R}^D \to \mathbb{R}$9, $\|x^i - x^j\| \ge d_{\min}$0.
• Stopping criteria: TolX = TolFun = TolFunHist = $\|x^i - x^j\| \ge d_{\min}$1, MaxIter = $\|x^i - x^j\| \ge d_{\min}$2, MaxFunEvals = $\|x^i - x^j\| \ge d_{\min}$3.
• Batch selection: The “clearing” method is preferred for computational efficiency after empirical comparison with Gurobi MILP and greedy heuristics.

4. Experimental Protocol and Baselines

Empirical evaluations use the 24 noiseless BBOB functions on $\|x^i - x^j\| \ge d_{\min}$4 for $\|x^i - x^j\| \ge d_{\min}$5, with budgets $\|x^i - x^j\| \ge d_{\min}$6 and $\|x^i - x^j\| \ge d_{\min}$7-batch extraction post hoc. Minimum distance thresholds vary by dimension:

• $\|x^i - x^j\| \ge d_{\min}$8: $\|x^i - x^j\| \ge d_{\min}$9
• $i \ne j$0: $i \ne j$1

The primary metric is the average loss: $i \ne j$2 where $i \ne j$3 is the known global minimum.

Comparative baselines include:

• Random uniform sampling
• Standard CMA-ES with restarts
• RS-CMSA (CMA-ES with repelling subpopulations)
• Hill-Valley EA
• WGraD clustering + DE

5. Results and Empirical Performance

Experimental results demonstrate that CMA-ES-DS consistently outperforms all baselines under medium to stringent distance constraints. Key findings include:

• Batch Extraction: The Gurobi MILP selector achieves the best average loss but is computationally expensive. The “clearing” method is 10–15% worse but runs in milliseconds, with negligible real-world impact on final solution quality.
• Distance Constraint Sensitivity: For weak constraints ($i \ne j$4 in 10D with $i \ne j$5), standard CMA-ES or RS-CMSA sometimes produces a better leader but frequently fails to assemble $i \ne j$6 valid, well-separated solutions for higher $i \ne j$7. Under strong constraints ($i \ne j$8 in 10D), CMA-ES-DS produces batches with 20–50% lower average loss than alternatives.
• Reliability: CMA-ES-DS yields $i \ne j$9-point batches in every trial, whereas other methods often fail for large $d_{\min}>0$0, evidenced by missing/dashed lines in reports.
• Variance: Variance across five independent runs is lowest for CMA-ES-DS, particularly in high-dimensional or evaluation-constrained conditions.

6. Strengths, Limitations, and Potential Extensions

Strengths:

• Guarantees the joint achievement of high-quality leadership and minimum-distance diversity in portfolio generation.
• Demonstrates robust performance across standardized black-box landscapes, a range of dimensionalities, and function evaluation budgets.
• Reliable in producing valid batches under all evaluated $d_{\min}>0$1 settings.

Limitations:

• Budget partitioning across $d_{\min}>0$2 instances can slightly reduce the leader’s quality when $d_{\min}>0$3 is very small, relative to dedicating the full budget to a single CMA-ES run.
• The approach natively assumes a Euclidean distance metric; adaptation is required for problem-specific metrics.

Potential Extensions:

• Adaptive, progress-dependent allocation of function evaluations among parallel instances.
• Extension to dynamic or anisotropic tabu regions (e.g., ellipsoidal neighborhoods) to enforce non-Euclidean diversity constraints.
• Integration of subset-selection heuristics into the training loop for active steering.
• Hybridization with Quality-Diversity paradigms to promote solution niching beyond geometric (input-space) distance alone.

CMA-ES-DS represents a pragmatic and conceptually straightforward method for generating high-quality, strictly diverse solution batches under user-defined input-space constraints, with demonstrated empirical superiority over previous state-of-the-art approaches (Santoni et al., 19 Feb 2025).