Research On Transfer Evolutionary Computing And Its Application

Evolutionary computation is a type of algorithm to solve optimization problems by simulating species evolution and clustering behavior in nature.It is widely used in black box optimization,combination optimization,non-convex optimization and multi-objective optimization.This paper focuses on a branch of evolutionary computing:Transfer Evolutionary Computing.Transfer evolutionary computing refers to a type of algorithm in which individuals in population assist in optimization through implicit or explicit transfer,and can be categorized into two types:Evolutionary Multitask Optimization and Parallel Evolutionary Computing.As an extension of traditional evolutionary computing,transfer evolutionary computing has broad application value and research significance.For evolutionary multitask optimization,there are two contributions in this paper:Firstly,for Multifactorial Evolutionary Algorithms(MFEA),the most classic algorithm in evolutionary multitask optimization,we theoretically analyze the inherent defects of MFEA in dealing with multitask problems of different dimensions,and proposed an improved version of Hetero-Dimensional Multitask Evolutionary Algorithms(HD-MFEA).In HD-MFEA,we propose hetero-dimensional assortative mating algorithm and self-adaption elite replacement algorithm,which can make HD-MFEA perform better genetic transfer in hetero-dimensional multitask problem.Meanwhile,we also propose benchmark problem for hetero-dimensional multitask optimization,in the test problem,HD-MFEA is better than MFEA and other improved algorithms.Secondly,we extend the application scope of evolutionary multitask optimization.For the first time,we equate the training problems of neural networks with different structures to the hetero-dimensional multitask problem.At the same time,in view of the hierarchical characteristics of neural networks,the hetero-dimensional multitask Neuroevolution algorithms(HD-MFEA Neuroevolution)is proposed,which can train multiple neural networks simultaneously.Through experiments on chaotic time series data sets,we found that the HD-MFEA Neuroevolution algorithm is far superior to other evolutionary algorithms,and its convergence speed and accuracy are better than the gradient algorithms commonly used in neural network training now.For parallel evolutionary computing,we take the lion swarm optimization algorithm as an example of evolutionary algorithm,and propose the parallel lion swarm algorithm for solving traveling salesman problem.Firstly,by introducing discrete coding and order crossover into the original lion swarm optimization algorithm,we propose the discrete lion swarm algorithm for solving combinatorial optimization problems;Then,we propose the parallel lion swarm algorithm based on the island model.The parallel lion swarm algorithm uses the ring topology structure and shares information by transferring the optimal individuals between adjacent subpopulations;Finally,we conducted experiments on different hardware platforms,verifying that parallel lion swarm algorithm can give a shorter route and achieve better acceleration performance on different platforms.