Improvement Of Genetic Algorithm With Surrogate Model

The genetic algorithm is a computational model introduced and studied by John Holland and his student DeJong in 1975.It draws on Darwin's theory of evolution,"the survival of the fittest,and the discomfort is eliminated".Using the fitness function to evaluate each individual in the group,through a certain evolutionary strategy to select and mutate,then new sample points(i.e.individuals)in the search space will be generated.Through continuous iteration,excellent individuals are saved and the global optimal solution is obtained.The genetic algorithm has no requirement for the continuity and the ductility of the objective function,and has good parallel processing capability and global search capability.There is no need to have prior knowledge of the problem.After random initialization,you can optimize the search space yourself,and then you can deal with problems such as minimizing or maximizing functions.Due to its good robustness,genetic algorithms are widely used in many scientific and industrial fields,such as robotics,automatic control,image processing,and neural networks.However,for some problems such as reservoir numerical fitting and dome structure optimization,these problems tend to optimize the high-dimensional parameters,require a large amount of fitness calculation(Expensive-computationally),and fitness evaluation.It can't be given by a specific formula,but by the result output of the auxiliary tool(Blackbox),we call this kind of problem a HEB problem.For such problems,although the genetic algorithm can obtain the global optimal solution,the computational process is very time consuming due to the need to calculate a large number of individual fitness,and the fitness calculation depends on the auxiliary tools,which leads to a large computational cost.This makes genetic algorithms inefficient in solving such problems.In order to solve this problem,this paper uses the surrogate model to predict the fitness value of the problem to be optimized,thereby greatly reducing the number of actual computational fitness.The backpropagation neural network,multiple linear regression and support vector regression are used to train the surrogate model,and the surrogate model is combined with the traditional genetic algorithm to obtain BPGA,MLRGA and SVRGA algorithms.Through the benchmark numerical experiments,the results show that the strategy can effectively reduce the number of calculations of actual fitness on the basis of obtaining a good global optimal solution.