Variational Genetic Algorithm Based On GPU Acceleration

As a meta-heuristic intelligent optimization algorithm characterized by population evolution,genetic algorithm has the advantages of global searching,efficient,parallel and self-learning without gradient data information,and is capable of adaptively finding optimal values and efficiently handling large-scale,complex nonlinear and discontinuous problems in engineering.However,real-life optimization problems often have a large number of parameters to be optimized,called "large-scale global optimization problems".Genetic algorithms require a large amount of computing power to find optimal solutions in the high-dimensional search space of these problems.Therefore,optimizing the search speed,efficiency,and quality of existing genetic algorithms is the basis for more widespread use of genetic algorithms.In this paper,the specific gene expression algorithm(SGEA)is proposed.Compared with existing genetic algorithms,this algorithm uses a two-level encoding model and introduces filtering operators that can weigh the development and exploration processes.The two-layer coding framework consists of a binary RNA coding layer and a decimal m RNA coding layer,which are connected by a fully connected neural network called the variational autoencoder,through which a uniformly distributed search space is mapped to a Gaussian distributed latent space.The filtering operator reconciles the genetic algorithm search time and search ability contradiction by calculating the outlier and fitness values of the individuals in the population,and making a comprehensive selection of the individuals in the population by roulette wheel.By using MANOVA,t-test and non-parametric Wilcoxon Ranksum test,this paper verifies that the variational genetic algorithm can significantly optimize the single-objective optimization genetic algorithm and multi-objective optimization genetic algorithm under statistical significance.And the improved genetic algorithm is used to solve two constrained single-objective engineering design problems to verify the overall feasibility of the algorithm.