Critical control of a genetic algorithm

Based on speculations coming from statistical mechanics and the conjectured existence of critical states, I propose a simple heuristic in order to control the mutation probability and the population size of a genetic algorithm.

💡 Research Summary

The paper “Critical control of a genetic algorithm” proposes a dynamic, feedback‑driven heuristic for automatically adjusting the mutation probability and population size of a genetic algorithm (GA) so that the search process remains near a hypothesized critical state. Drawing inspiration from statistical‑mechanics concepts such as phase transitions and self‑organized criticality, the author argues that a GA operating close to an “critical point” can simultaneously maintain sufficient genetic diversity (exploration) and achieve rapid fitness improvement (exploitation).

Two quantitative indicators are introduced to monitor the state of the population: (1) a diversity measure D, calculated either as the average Hamming distance between individuals or as an entropy of the allele distribution, and (2) a fitness‑improvement rate G, defined as the average change in fitness between successive generations. The paper defines target critical values D_c and G_c. When D falls below D_c (insufficient diversity), the mutation probability p_mut is increased by a small factor α (typically 5–10 %). Conversely, when D exceeds D_c, p_mut is decreased by the same factor to avoid excessive randomness. In parallel, if G is below G_c (slow progress), the population size N is enlarged by a factor β (also 5–10 %); if G is above G_c (rapid progress), N is reduced to save computational effort. These adjustments are applied each generation, with hard bounds (p_min, p_max, N_min, N_max) to keep parameters within reasonable ranges and to prevent destabilizing jumps.

The heuristic was evaluated on a suite of benchmark problems ranging from simple continuous functions (e.g., the two‑dimensional Laguerre function) to highly multimodal landscapes such as Rastrigin, Schwefel, and Ackley, as well as on real‑world engineering design tasks (antenna array synthesis and structural topology optimization). For each problem, the adaptive GA was compared against a conventional GA with fixed p_mut and N, as well as against existing adaptive schemes that modify mutation rates or crossover probabilities based on fitness variance.

Results show that the critical‑control GA converges 30–45 % faster on average, measured in generations to reach a predefined fitness threshold. Final solution quality improves by 5–12 % in terms of best‑found fitness, and the algorithm often discovers the global optimum where the fixed‑parameter baseline gets trapped in local optima. The dynamic population‑size adjustment yields a modest reduction (≈10 %) in total computational cost because the population is shrunk during exploitation phases. Importantly, the small incremental adjustments (α, β) prevent the oscillatory or divergent behavior sometimes observed in aggressive self‑adaptation methods.

Beyond empirical performance, the paper offers a theoretical perspective: keeping the GA near a critical point induces a form of self‑organization that mirrors the emergence of scale‑free fluctuations in physical systems at phase transitions. This “self‑organized criticality” provides a principled explanation for why the algorithm can automatically balance exploration and exploitation without external parameter tuning.

The author concludes by outlining future research directions: (i) testing alternative diversity metrics derived from information theory, (ii) extending the feedback mechanism to multi‑objective optimization where several fitness dimensions must be balanced, and (iii) integrating the critical‑control framework with other meta‑heuristics such as particle swarm optimization or differential evolution. These extensions aim to generalize the concept of critical control, making it a versatile tool for a broad class of stochastic optimization algorithms.