Linkage-Guided Genetic Variation: Overcoming Operator Blindness in Genetic Algorithms

ICLR 2026 Conference Submission14206 Authors

Keywords: Genetic Algorithms; Building Blocks; Linkage Disequilibrium; Neural Architecture Search; Multi-Objective Optimization

Abstract: The core bottleneck of Genetic Algorithms is operator blindness: crossover and mutation locations are chosen at random, routinely breaking valuable building blocks. We introduce the Evolving Locus Linkage Graph (ELLG), which embeds the linkage principle (keep strong segments intact, recombine at weak boundaries) into operator design. At each generation, ELLG updates per-locus linkage weights from observed fitness, producing a task-specific linkage map that tells the algorithm which segments to keep intact and where to cut; as generations proceed, these protected regions and preferred cut sites become increasingly well-defined. A simple monotone transformation converts the learned weights into placement probabilities for crossover and mutation, replacing uniform randomness with targeted, structure-aware operator placement. We integrate ELLG as a plug-in to a standard GA without changing the problem encoding or operator semantics. We benchmark ELLG against a large pool of state-of-the-art evolutionary methods across two domains: classical multi-objective optimization suites and Neural Architecture Search, ELLG achieves higher final solution quality in experiments.

Primary Area: optimization

Submission Number: 14206