ELA-Space Fitness-Sharing Mechanism

• The ELA-space fitness-sharing mechanism is a method that uses Exploratory Landscape Analysis features to penalize candidate functions with similar landscapes, promoting underrepresented regions.
• It integrates an adaptive niche radius, computed via the Manhattan distance between standardized ELA feature vectors, to dynamically adjust shared fitness during evolutionary loops.
• Empirical evaluations show significantly increased diversity with higher median nearest neighbor distances, confirming its effectiveness in mitigating redundancy in generated optimization landscapes.

The ELA-space fitness-sharing mechanism is an approach introduced to increase population diversity in LLM-driven design loops for continuous optimization problems, specifically to address the structural redundancy of generated benchmark test suites. By leveraging Exploratory Landscape Analysis (ELA) features as a representation of problem space, this mechanism penalizes candidate functions whose landscape characteristics are overly similar, thereby guiding the search toward underrepresented regions in ELA feature space and expanding the diversity of generated optimization landscapes (Skvorc et al., 26 Jan 2026).

1. Formal Mechanism and Mathematical Foundation

Fitness sharing in ELA space operates by adjusting the raw fitness scores—typically the property-prediction output of a trained model (e.g., an XGBoost predictor)—based on the similarity between candidates as measured in landscape feature space. Each candidate function $f_i$ is encoded by an ELA feature vector: $$\bar\phi_{f_i} = \mathbb{E}_{X\sim p_X}\bigl[\mathrm{ELA}(X,f_i(X))\bigr] \in \mathbb{R}^m$$ where $p_X$ is the input distribution and $m$ is the number of ELA features. All feature vectors are $z$-standardized before distance calculations.

Similarity between two candidates $f_i$ and $f_j$ is computed using the $L_1$ (Manhattan) norm: $$D_{\mathrm{ELA}}(f_i,f_j) = \|\bar\phi_{f_i} - \bar\phi_{f_j}\|_1$$

Fitness sharing applies a linear kernel: $$\mathrm{sh}(d) = \max\bigl(0,\;1 - (d/\sigma)\bigr)$$ where $\sigma$ is the adaptive niche radius set to the mean pairwise $L_1$ distance among all candidates: $$\sigma = \frac{2}{N(N-1)} \sum_{1\le i<j\le N} D_{\mathrm{ELA}}(f_i,f_j)$$ with $N$ the number of candidates in the population (or offspring set per generation).

The final "shared" fitness for $f_i$ is: $$\hat F_i = \frac{F_i}{\displaystyle\sum_{j\neq i} \mathrm{sh}\bigl(D_{\mathrm{ELA}}(f_i,f_j)\bigr)}$$ A crowded niche decreases $\hat F_i$, favoring individuals in sparsely populated feature regions.

2. Algorithmic Integration in Evolutionary Loops

The ELA-space fitness-sharing mechanism is integrated into an evolutionary LLM-problem generation framework as follows:

1. Candidate Generation: A LLM mutates and recombines problem descriptions ("prompts") and generates new function code (offspring).
2. Feature Evaluation: For each generated function, features are computed via ELA, standardized, and raw fitness is estimated using a property-predictor model.
3. Distance and Sharing Calculation: All pairwise ELA distances are computed, $\sigma$ is recalculated each generation, and sharing penalties are applied.
4. Selection: Shared fitnesses $\hat F_i$ are used to rank candidates, selecting the top-$\mu$ for the next parent pool.
5. Termination: The process iterates over a fixed number of generations, after which outputs are optionally post-filtered using more expensive verification procedures such as basin-counting or separability testing.

The following table summarizes the main inputs and processes per generation:

Step	Description	Key Output
Generation	LLM mutation/recombination of prompts	Offspring function code
Evaluation	ELA computation, z-scoring, raw fitness prediction	Feature vectors, $F_i$
Sharing	Distance calculation, $\sigma$ update, $\hat F_i$ computation	Shared fitnesses
Selection	Rank and select by $\hat F_i$	Next-generation parents

3. Parameterization and Tuning

The mechanism's principal parameters and their respective rationale are as follows:

• Population sizes ($\mu$, $\lambda$): Typically set to small or medium values (e.g., $\mu=8$, $\lambda=16$) to manage computational burden while enabling meaningful exploration.
• Niche radius ($\sigma$): Dynamically adapted each generation via the mean pairwise distance. This self-scaling is essential to avoid manual tuning across heterogeneous feature scales.
• Sharing exponent ($\alpha$): Fixed at 1 (linear decay in the kernel), as empirical testing found this effective without need for further adjustment. Larger $\alpha$ would sharpen penalization for near-duplicate landscapes.
• Distance norm: Manhattan ($L_1$), chosen for robustness to outliers and the interpretability of per-feature deviations.

No evidence in the cited work indicated a need for elaborate hyperparameter search in practice.

4. Empirical Impact on Diversity

Ablation studies were conducted using both an open-source model (qwen2.5-coder_14b) and a proprietary model (gpt5-nano) with and without the ELA-space fitness sharing mechanism, targeting simultaneous basin-size homogeneity and separability.

• Without sharing: Nearest-neighbor (NN) distances in ELA feature space among generated functions concentrated at low values ($\sim$2.0–2.5).
• With sharing: NN distance distribution shifted significantly higher ($\sim$3.0–4.0), indicating that generated functions occupy more distant regions in ELA space.
• Model-specific result: For gpt5-nano, the median NN distance increased from $\sim$2.2 (no sharing) to $\sim$4.1 (with sharing).
• Significance: Two-tailed Mann–Whitney U-tests confirmed the observed shifts were highly significant ($p \ll 0.01$).
• Visualization: t-SNE embeddings demonstrated that shared fitness pushes generations into BBOB gaps rather than clustering around known test function families.

This suggests that ELA-space sharing is effective in mitigating landscape redundancy and systematically broadening the spread of generated benchmarks (Skvorc et al., 26 Jan 2026).

5. Adaptation and Application in Broader Contexts

The ELA-space fitness-sharing mechanism generalizes to any domain involving automated generation of problem instances characterized by behavioral or feature descriptors. Essential steps include:

• Substitute Features: ELA features may be replaced by any domain-appropriate feature map.
• Auto-adaptive Niche Radius: Using an adaptive $\sigma$ based on mean pairwise distance addresses issues of feature scaling and obviates manual tuning.
• Tunable Kernel: The sharing exponent $\alpha$ controls the aggressiveness of near-duplicate penalization.
• Algorithm Selection: The method directly supports diversity–quality trade-offs in evolutionary frameworks, and can be combined with property-targeting objectives or more general behavior-diversity paradigms such as novelty search.

In continuous black-box optimization, adopting ELA-space sharing within the generation loop prevents overproduction of near-identical problem instances and achieves wider, more interpretable coverage of landscape structure. A plausible implication is that similar sharing-based approaches can systematically mitigate redundancy in algorithm configuration, reinforcement learning environments, or any benchmark suite design where landscape feature novelty is desired.

6. Summary and Outlook

The ELA-space fitness-sharing mechanism constitutes a lightweight yet principled means to enforce diversity in the LLM-driven evolution of continuous optimization problems. By penalizing population crowding in feature space through shared-fitness adjustment, it specifically addresses the "redundant landscape" problem prevalent in automated problem design. The mechanism has demonstrated statistically significant impacts on the spread of generated instances in ELA space and provides a template for generalization to any evolutionary search or generation framework where feature-based diversity is a core requirement (Skvorc et al., 26 Jan 2026).