EventsThe 3rd International Online Conference on Biomimetics
Published
This submission belongs to the session S5. Biologically Inspired Optimisation and Management of the event The 3rd International Online Conference on Biomimetics
Published date
19 Nov, 2026
Academic Editor
author-avatarJulian Vincent
Contribution&Funding
Conceptualization,
[George Novikov]
Methodology,
[George Novikov]
Software,
[George Novikov]
Validation,
[George Novikov]
Formal analysis,
[George Novikov]
Investigation,
[George Novikov]
Data curation,
[George Novikov]
Writing - original draft,
[George Novikov]
Funding: --
Citation
George Novikov, Sergey Dolenko, Igor Isaev, George Alexandrovich Novikov, Age-Structured Genetic Algorithms for Robust Convergence and Diversity Preservation, in Proceedings of The 3rd International Online Conference on Biomimetics, 24 November–26 November 2026, MDPI: Basel, Switzerland
Share
Email
Facebook
Twitter
LinkedIn

Age-Structured Genetic Algorithms for Robust Convergence and Diversity Preservation

Igor Isaev 2
1. Physics, Moscow State University named after M.V. Lomonosov, Moscow, 119234, Russia
2. Physics, D.V. Skobeltsyn Institute of Nuclear Physics, Russia
Abstract

Introduction

Gender-based genetic algorithms (GAs) have previously been studied as a bioinspired extension of the classical GA. In the present work, the main new mechanism is age-based population control. We investigate whether limiting the lifetime of long-surviving individuals, applying age-dependent selection pressure, and protecting a small elite can reduce premature convergence. Here, premature convergence means that the search reaches a stable plateau too early and stops producing substantial improvements, while other variants operating under the same computational budget can still find better solutions.

The age mechanism periodically renews long-surviving individuals and is intended to limit prolonged dominance of the same solutions. Genetic diversity was not measured directly; therefore, diversity preservation is treated as a plausible mechanism rather than a directly demonstrated result. A complementary Hamming-distance selection framework for binary chromosomes is being developed to introduce explicit control of genetic differences during parent selection.

Methods

Ten binary-coded GA variants were compared, including the classical GA, previously studied gender-based variants, and six schemes incorporating age-based mechanisms. GGA-MM is a previously proposed gender-based GA in which two groups of individuals use different bit-mutation probability profiles. Here it serves as the reference architecture for the age mechanisms and is not the principal novelty.

The six age-based schemes include the classical GA with a hard lifespan of 15 generations; GGA-MM with hard lifespans of 7, 15, or 25 generations; GGA-MM with soft aging after age 10; and GGA-MM with a hard lifespan of 15 generations in which the three best individuals are protected from age-based removal. Each newly generated offspring receives age 0, so age is not inherited. At the beginning of each generation, the age of surviving individuals is increased by one. Under hard aging, an individual reaching the maximum age is removed and replaced by a new random individual with age 0. In practice, age accumulates mainly in elite individuals because they are transferred directly between generations. Under soft aging, an age-dependent penalty is applied to the effective objective-function value after the threshold, reducing the probability of further selection without forced removal.

Eight standard ten-dimensional benchmark functions—Ackley, Griewank, Levy, Rastrigin, Rosenbrock, Schwefel, Sphere, and Zakharov—were encoded using 10 bits per variable. They were used to test whether the effect of age-based control extends beyond spectral decomposition to different multidimensional optimization landscapes. Each algorithm used 150 individuals for 100 generations and was evaluated in 30 independent runs per function, giving 2,400 benchmark runs. All variants had the same computational budget of 15,000 objective-function evaluations per run.

Performance was assessed using the final objective-function value, interquartile range (IQR), and convergence epoch. Convergence was defined operationally as a practical plateau: if no improvement greater than 10^-6 was observed for 10 consecutive generations, the epoch preceding this stagnation interval was recorded as the convergence epoch. This measure characterizes search stabilization and does not imply that a local or global optimum has been reached. Pairwise comparisons with the classical GA used two-sided Mann–Whitney tests with Holm correction.

Ten noisy spectra composed of four Gaussian components, with the leading-band separation reduced from 350 to 90 cm^-1, provided an additional physically interpretable benchmark. The spectral study comprised 3,000 runs, giving 5,400 independent runs across the two experimental blocks. Increasing band overlap eventually causes loss of reliable identifiability of individual component parameters.

For binary vectors x and y, the Hamming distance is the number of bit positions at which they differ. In the developing diversity-aware framework, we compare one fitness-selected and one diversity-selected parent, probabilistic mixtures of both criteria, and selection of both parents using Hamming separation. Whole-chromosome and gene-wise distances are also compared.

Results

Across the eight benchmark functions, age-controlled GGA-MM variants generally achieved lower median final objective-function values than the classical GA. Hard-age-25 and immune-elite were significantly better than the classical GA in all eight Holm-adjusted comparisons and obtained mean fitness ranks of 3.250 and 3.125, respectively, versus 8.750 for the classical GA. Lower rank corresponds to better overall solution quality.

Hard-age-25 reduced the IQR of final objective-function values for every benchmark function. Its median IQR ratio relative to the classical GA was 0.148, meaning that its typical IQR was about 15% of that of the classical GA. This indicates greater reproducibility across independent runs, but not direct evidence of higher genetic diversity within the population.

The classical GA frequently reached a plateau earlier but at substantially worse final objective-function values. Therefore, early stabilization should not be interpreted as an unconditional speed advantage; rather, the observed pattern is consistent with premature stabilization around a poorer solution.

On the spectral-decomposition problems, hard-age-25 achieved the best aggregate objective-function rank and outperformed the classical GA on all ten spectra. However, no modification eliminated the loss of parameter identifiability near the physical resolution limit. Improving the evolutionary search is therefore useful while the spectral components remain physically distinguishable, but cannot compensate for information loss caused by excessive overlap.

Conclusions

The main contribution is age-based control of individual lifetime in a genetic algorithm. The results show that early stabilization is not necessarily evidence of efficient search: the classical GA often reached a plateau earlier but ended with poorer objective-function values.

The most consistent results were obtained when strong solutions were allowed to participate in evolution for a sufficiently long time while indefinite dominance was restricted. Hard aging with a lifespan of 25 generations and the immune-elite scheme provided the best balance between retaining strong solutions and periodically renewing the population.

We do not claim that increased genetic diversity was directly demonstrated. Direct confirmation would require explicit measurement of differences among individuals within the evolving population. The developing Hamming-distance approach is the next step in this direction because it introduces an explicit measure of genetic difference into parent selection. Alternative selection regimes and whole-chromosome versus gene-wise distance representations form a separate part of the conference presentation.

Keywords
genetic algorithm; age-structured evolution; Hamming distance; diversity-aware selection; premature convergence; multimodal optimization; biomimetic computing
Fabrication and Thermal Testing of an Elephant Pinna-Inspired Heat Exchanger for Closed-Loop Data Centre Cooling
Biomimetic Engineering of Mycelial Pulp: A Sustainable Manufacturing Strategy for Wood Cellulose Alternatives