[Paper] Discrete Gene Crossover Accelerates Solution Discovery in Quality-Diversity Algorithms

Source: arXiv - 2602.13730v1

Overview

The paper introduces a discrete gene‑level crossover operator for Quality‑Diversity (QD) algorithms. By borrowing the idea of biological meiosis, the authors show that swapping whole genes between elite solutions can dramatically speed up the discovery of high‑performing, diverse behaviours—especially in later stages of optimisation where traditional mutation stalls.

Key Contributions

• Novel mutation operator that combines standard variation (e.g., Gaussian mutation) with a gene‑wise crossover mechanism.
• Theoretical justification linking the operator to biological recombination, highlighting its ability to move whole “building blocks” across the population.
• Empirical evaluation on three locomotion benchmarks (e.g., planar walkers, quadrupeds) demonstrating consistent gains in QD score, coverage, and peak fitness.
• Analysis of dynamics showing that the crossover shines after a useful archive of elites has been built, helping the algorithm escape stagnation.
• Open‑source implementation (released with the paper) that can be dropped into existing QD frameworks such as MAP‑Elites or Novelty Search.

Methodology

1. Baseline QD framework – The authors start from a standard MAP‑Elites pipeline: an archive indexed by behavioural descriptors, elite individuals stored per niche, and a variation loop that samples elites, mutates them, and attempts to insert offspring.
2. Gene‑level crossover mutation – When generating an offspring, they first pick two elite parents. For each gene (parameter) in the genotype, they flip a biased coin: with probability p they copy the gene from parent A, otherwise from parent B. This yields a discrete recombination rather than a blended interpolation.
3. Hybrid variation – After crossover, the offspring undergoes a conventional mutation (e.g., Gaussian noise) to retain fine‑grained exploration.
4. Experimental setup – Three simulated locomotion tasks (2‑D biped, 3‑D quadruped, and a modular snake) are used. Each run is repeated 10 times, and performance is measured by:
   • QD Score (sum of fitnesses across the archive)
   • Coverage (fraction of behavioural niches filled)
   • Max Fitness (best individual found)
5. Statistical analysis – Results are compared against a pure‑mutation baseline using Mann‑Whitney U‑tests and effect‑size metrics.

Results & Findings

• Early‑stage parity – In the first few thousand evaluations, both methods perform similarly; the crossover does not hinder initial exploration.
• Late‑stage boost – After ~10⁴ evaluations, the crossover‑augmented runs start to outpace the baseline, indicating that the operator excels at recombining already discovered “building blocks.”
• Robustness across tasks – All three environments show the same trend, suggesting the approach is not task‑specific.
• Statistical significance – Improvements are significant (p < 0.01) with medium to large effect sizes.

Practical Implications

• Faster prototyping of diverse behaviours – Engineers using QD for robot gait synthesis, game content generation, or neural architecture search can reach richer solution sets with fewer simulation hours.
• Plug‑and‑play upgrade – The crossover operator is a thin wrapper around existing elite‑selection code; integrating it into libraries like pymap_elites or QDpy requires only a few lines of configuration.
• Better utilisation of compute budgets – In cloud‑based optimisation pipelines, the extra performance translates directly into cost savings (e.g., 20 % fewer GPU‑hours for the same archive quality).
• Potential for hybrid AI pipelines – Combining discrete crossover with gradient‑based fine‑tuning could yield a “coarse‑to‑fine” search strategy, useful for domains where gradient information is noisy or unavailable.

Limitations & Future Work

• Gene independence assumption – The crossover treats each gene independently; highly epistatic (inter‑dependent) parameters may not benefit as much.
• Scalability to very high‑dimensional genotypes – Experiments capped at ~200 parameters; performance on thousands‑dimensional neural nets remains untested.
• Dynamic crossover probability – The paper uses a fixed per‑gene swap probability; adaptive schemes (e.g., based on archive diversity) could further improve results.
• Real‑world validation – All tests are in simulation; transferring the approach to physical robots or hardware‑in‑the‑loop optimisation is an open next step.

Bottom line: Adding a simple, biologically inspired gene‑level crossover to QD algorithms gives developers a low‑cost, high‑impact tool for accelerating the discovery of diverse, high‑quality solutions—especially when the search has already uncovered useful building blocks.

Authors

• Joshua Hutchinson
• J. Michael Herrmann
• Simón C. Smith

Paper Information

• arXiv ID: 2602.13730v1
• Categories: cs.NE
• Published: February 14, 2026
• PDF: Download PDF