Geno-Synthetic Algorithm hacks heterogeneous optimization with type-native operators
GSA handles complex-valued descriptors and embedding vectors where other methods fail completely.
Standard evolutionary algorithms force heterogeneous parameters—integers, reals, Booleans, categoricals, complex numbers, embedding vectors—into a single chromosome, then apply generic operators with rounding and repair, losing representational fidelity. Alex Bogdan’s new Geno-Synthetic Algorithm (GSA) solves this by partitioning gene families by representational type, evolving each in parallel with type-native operators, then assembling them into executable phenotypes for joint fitness evaluation. The framework is formalized as a typed product-space search with an explicit assembly operator, and an MIT-licensed open-source implementation is released.
In an empirical study across seven benchmark problems (six synthetic plus COCO BBOB-MixInt) at budgets from 5,000 to 100,000 evaluations, GSA proves uniquely capable of handling complex-valued descriptors and embedding vectors—where all baselines fail. On smooth synthetic multi-family problems, well-tuned flattened differential evolution remains strongest; but on BBOB-MixInt at 100k evaluations, GSA_DIRECT becomes statistically indistinguishable from FLATTENED_DE while FLATTENED_EA drops from second to fifth rank. Ablations confirm type-native operators are essential, elite credit dominates ensemble credit, and active assembly outperforms passive concatenation. The framework extends naturally to prompt and embedding optimization for large language model systems, making it directly relevant for AI engineers.
- GSA partitions gene families by type (integer, real, Boolean, categorical, complex, embedding) and evolves them with type-native operators rather than generic ones.
- In benchmarks, GSA is the only method that successfully handles complex-valued descriptors and embedding vectors—critical for modern AI pipelines.
- On BBOB-MixInt at 100k evaluations, GSA_DIRECT matches the performance of flattened differential evolution, while traditional flattened evolutionary algorithms degrade significantly.
Why It Matters
GSA enables more faithful, efficient optimization for composite real-world problems and LLM prompt/embedding tuning.