Biomimetic use of genetic algorithms

Genetic algorithms are considered as an original way to solve problems, probably because of their generality and of their “blind” nature. But GAs are also unusual since the features of many implementations (among all that could be thought of) are principally led by the biological metaphor, while efficiency measurements intervene only afterwards. We propose here to examine the relevance of these biomimetic aspects, by pointing out some fundamental similarities and divergences between GAs and the genome of living beings shaped by natural selection. One of the main differences comes from the fact that GAs rely principally on the so-called implicit parallelism, while giving to the mutation/selection mechanism the second role. Such differences could suggest new ways of employing GAs on complex problems, using complex codings and starting from nearly homogeneous populations.

💡 Research Summary

The paper critically examines the relationship between genetic algorithms (GAs) and natural evolution, arguing that most GA implementations rely heavily on a biological metaphor while efficiency considerations are added only after the fact. The author first highlights a fundamental similarity: in both systems selection does not act on whole individuals but on sub‑units—schemata in GAs and genes in biology. Holland’s concept of implicit parallelism, which counts the number of short schemata simultaneously present in a random population, is shown to be valid only for initially random populations. In nature, however, selection truly operates on genes, and recombination (crossover) is a ubiquitous process, even among unicellular organisms.

The paper then identifies two conflicting constraints that any GA designer must balance. Complexity: For non‑separable problems the mapping from genotype to phenotype must be intricate; simple binary encodings split the problem into independent sub‑problems, which is inappropriate when fitness contributions are highly epistatic. High‑order schemata—genes that modify the evaluation of other genes—are proposed as a way to achieve this complexity. Continuity: Small mutations should not cause large fitness swings. Natural genes are “extended” (coding for proteins) so that a single amino‑acid change often has a modest effect. Conventional GA encodings concentrate too much meaning on each bit, violating continuity. An extended coding where mutations have limited phenotypic impact is therefore required.

Macro‑operators such as inversion, translocation and duplication are acknowledged as biologically plausible, but the paper warns against a complete separation of gene and locus. In DNA, signals (promoters, terminators) decouple absolute position from function, allowing genes to move without loss of meaning. GA designs should emulate this signal‑based approach rather than a rigid gene‑locus mapping.

A central, perhaps counter‑intuitive, proposal is to start GA runs with nearly homogeneous populations. In nature, individuals that successfully recombine share >99 % of their genome, and most differences are neutral. If a GA uses a complex genotype‑phenotype mapping, a random initial population would produce offspring with little or no fitness because crossover would generate highly disruptive combinations. By contrast, a homogeneous starting point ensures that crossover makes only minor changes, allowing the mutation‑selection mechanism to dominate the search. This shifts the algorithm’s driving force away from implicit parallelism toward a gradual, continuous walk through the fitness landscape, which the author argues is more “biomimetic”.

The paper illustrates these ideas with two case studies. Lavalou (1990) encoded short musical pieces using recursive operators and let a connectionist network evaluate them, demonstrating that a complex coding can generate rich phenotypes. Werner & Dyer (1992) evolved agents whose genomes encoded small neural controllers, showing how genotype‑phenotype complexity can be increased further. Both examples support the claim that traditional GA encodings are too limited for highly multidimensional, non‑separable problems.

In the concluding section the author summarizes the defining features of a biomimetic GA: (1) high‑order, combinatorial genotype‑phenotype mapping; (2) extended genes to preserve continuity; (3) signal‑based macro‑operators rather than strict gene‑locus separation; and (4) nearly homogeneous initial populations where mutation and selection dominate. While acknowledging that such algorithms may be slower and less efficient than classic GAs that exploit implicit parallelism, the paper argues that they open a new avenue for tackling problems that resist conventional decomposition and that may benefit from a more faithful imitation of natural evolutionary dynamics.