• Konstantin S. Privalov, Financial University under the Government of the Russian Federation (Moscow, Russia)

The relevance of this study is determined by the fact that the efficiency of classical genetic algorithms (GAs) in solving global optimization problems significantly depends on the choice of crossover and mutation probabilities, while fixed parameter values often lead to premature convergence and stabilization of the population in the vicinity of local extrema. The aim of this work is to develop and experimentally evaluate approaches to adaptive control of GA parameters based on artificial neural networks and reinforcement learning. Within a unified mathematical formulation, population state features, a set of actions represented by discrete changes in mutation and crossover probabilities p_m and p_c, and a reward function reflecting the improvement in solution quality between generations with a penalty for excessively high mutation are defined. Four algorithmic variants are considered: a classical GA with fixed parameters, a hybrid GA with a neural-network controller (GA+NN), a GA with tabular Q-learning (GA+RL), and the proposed parameter control method based on deep Q-learning, which uses a neural network to approximate the Q-function (GA+DQN). The scientific novelty of the study lies in the integration of a DQN agent into the GA parameter control loop within a formalized “state–action–reward” model and in comparing its efficiency with that of a neural-network controller and tabular Q-learning on continuous optimization problems. Numerical experiments were carried out on the Rastrigin and Schaffer test functions with 20 independent runs for each configuration. The final metrics used were the best objective function value in the last generation f_min(T_max) and the best value achieved over the entire run. It is shown that GA+RL provides the greatest improvement in solution quality. The GA+DQN method demonstrates a moderate improvement over the baseline GA, confirming the applicability of deep Q-function approximation to parameter control. The neural-network controller in the considered training scheme shows high sensitivity to parameter settings and, in these experiments, performs worse than the reinforcement learning approaches. The comparison results are presented in the form of convergence plots, an analysis of population diversity indicators, and a summary table.

genetic algorithm, adaptive parameter control, reinforcement learning, hybrid evolutionary algorithms

2026-06-05