The Geno-Synthetic Algorithm: Type-Factored Coevolutionary Optimization for Heterogeneous Genotypes and Assembled Phenotypes

Problem
This paper addresses the limitations of standard evolutionary algorithms (EAs) in optimizing heterogeneous design objects, which consist of diverse parameter types such as integers, real values, Booleans, categoricals, complex-valued descriptors, and embedding vectors. Traditional EAs flatten these heterogeneous parameters into a single chromosome, leading to a loss of representational fidelity and suboptimal performance. The authors propose the Geno-Synthetic Algorithm (GSA) as a solution to this gap. This work is presented as a preprint and has not yet undergone peer review.

Method
The Geno-Synthetic Algorithm introduces a type-factored coevolutionary framework that partitions gene families based on their representational types. Each gene family is evolved in parallel using type-native operators, which are specifically designed for the characteristics of the data type. The GSA is formalized as a typed product-space search procedure, incorporating an explicit assembly operator that constructs executable phenotypes for joint fitness evaluation. The authors provide an open-source reference implementation (gsa-experiments, MIT-licensed) to facilitate reproducibility. The empirical study evaluates eight GSA variants against five baseline algorithms across seven benchmark problems, including six synthetic problems and the COCO BBOB-MixInt suite, with evaluation budgets ranging from 5,000 to 100,000 evaluations.

Results
The GSA demonstrates significant architectural advantages, particularly in scenarios involving complex-valued descriptors or embedding vectors, where it is the only method capable of effective operation. In the context of smooth synthetic multi-family problems, a well-tuned flattened differential evolution algorithm remains the strongest baseline. However, on the BBOB-MixInt benchmark at 100,000 evaluations, the GSA variant GSA_DIRECT achieves performance that is statistically indistinguishable from the best-performing flattened differential evolution method (FLATTENED_DE), while the performance of FLATTENED_EA declines from second to fifth rank, indicating an asymptotic crossover. Ablation studies reveal that type-native operators are crucial for performance, elite credit assignment is more effective than ensemble credit, and active assembly outperforms passive concatenation on gated benchmarks.

Limitations
The authors acknowledge that their study is limited to specific benchmark problems and may not generalize to all optimization scenarios. They do not address the computational overhead associated with the type-factored approach, which may be a concern in high-dimensional or real-time applications. Additionally, the performance of GSA in highly dynamic environments or with noisy fitness landscapes remains unexplored.

Why it matters
The introduction of the Geno-Synthetic Algorithm has significant implications for the field of evolutionary computation, particularly in optimizing complex systems with heterogeneous parameters. By maintaining representational fidelity through type-factored evolution, GSA opens avenues for more effective optimization in various domains, including engineering design, machine learning model tuning, and large language model prompt optimization. The framework’s adaptability to complex data types positions it as a valuable tool for future research and practical applications in coevolutionary optimization.

Authors: Alex Bogdan
Source: arXiv:2605.13365
URL: https://arxiv.org/abs/2605.13365v1