Reference library
Neural network crossover in genetic algorithms using genetic programming
Paper summary
Genetic programming evolves expressions that combine parental weight vectors. Pair-specific disposable operators with a rising crossover rate improve final validation cross-entropy relative to no crossover on five classification datasets in 30-run comparisons. Simple mean crossover has significant benefits on only two datasets. Initial fixed-rate screening favors reusable operators, while the later schedule favors disposable ones; reusable operators converge to arithmetic means. Mutation uses gradient descent and pair-specific crossover adds inner search/evaluation work.
Relevance to the project
Assistant interpretation: useful evidence that operator and schedule choice matter, and that crossover can help despite its reputation. Pair-specific search could be considered only after explicit genotype/budget decisions; its classifier losses do not define ecological fitness or approve a new optimizer.
Five classification datasets, hybrid gradient learning, tuned schedule and extra GP evaluations; no equal-total-compute superiority established. Results are not a universal cut-point recipe or evidence that independently useful modules remain compatible.
Source record
- Reference ID
- neural-crossover-genetic-programming-2024
- Archived version
- Publisher version of record, 21 February 2024, 30-page PDF, CC BY4.0
- Summary provenance
- reference-library/papers/neural-crossover-genetic-programming-2024/metadata.json · summary
- Review status
- Targeted primary full-text review; not exhaustive reading or reproduction
- Reading scope
- Publisher HTML sections3–6; targeted original PDF pp.6–7,14–16,18–28 covering fitness, holdout split, extra search, parameters, all three comparisons and Tables3–10. Figure captions read; figures not exhaustively visually audited. No reproduction/code audit.