Multi-task Optimization Based On Evolutionary Algorithm

Multi-task optimization(MTO)is an emerging topic in the field of optimization.The goal of MTO investigates is to solve multiple optimization problems simultaneously and effectively.Evolutionary multi-task optimization introduces evolutionary algorithm into multi-task optimization,and performs across-domain optimization of multiple optimization problems by using the implicit parallelism of population-based search.Compared with traditional single-task optimization,evolutionary multi-task optimization enables the knowledge transfer between different optimization problems.The sharing of potential similarities and complementarities among optimization problems promotes the evolution of the populations,thereby leading to the improved performance and efficiency of problem-solving.Based on this framework,a multifactorial evolutionary algorithm(MFEA)has been proposed and demonstrated excellent performance on many multi-task optimization problems.In this dissertation,three improved MFEAs are proposed as follows:1)An adaptive memetic algorithm(AMA)is proposed,to improve the MFEA from three aspects.Firstly,a local search strategy based on individual transfer is introduced to enhance the learning efficiency of tasks.Secondly,a re-initialization technique is presented to update the individuals that are difficult to evolve.Finally,an adaptive selection strategy of parent individuals is adopted catering to the requirements of different stages of population evolution.The experimental results show that the overall performance of AMA is better than the MFEA.However,it tends to be susceptible to negative transfer.2)A two-level transfer learning algorithm(TLTLA)is proposed to deal with the negative transfer issue while improving the search efficiency.The upper level of TLTLA performs inter-task knowledge transfer.Considering the strong randomness of knowledge transfer caused by chromosomal crossover across different tasks,a learning strategy based on elite individual transfer is introduced to promote the search efficiency.The lower level of TLTLA performs intra-task knowledge.To alleviate the interference of negative transfer on population evolution,a one-dimensional search strategy based on unified coding representation is presented to deeply optimize individuals.Numerical experiments show that TLTLA has excellent performance on nine sets of two-task optimization problems.The experiments on six sets of three-task optimization problems show the high scalability of TLTLA.3)A multiple-search multi-objective multifactorial algorithm(MS-MOMFEA)is proposed to extend MFEA to handle multi-objective optimization problems.MS-MOMFEA is built on the multi-objective multifactorial evolutionary algorithm,and featured with a cross-dimensional decision variable search strategy and a prediction-based individual search strategy.The cross-dimensional decision variable search strategy uses the genetic information from multiple dimensions to optimize the decision variables,which ensures high quality solutions are generated by transferring useful information and discarding negative information.The prediction-based individual search strategy estimates a possible population center based on historical data.Individual mapping around the predicted population center is performed to maintain the diversity in the late stage of population evolution.Compared with the classic NSGA-II algorithm and multi-objective multifactorial evolutionary algorithm,MS-MOMFEA shows better convergence performance and is capable of finding a set of solutions closer to the real Pareto front.This dissertation proposes new algorithms to solve single-objective and multi-objective multi-task optimization problems.The empirical studies have demonstrated the effectiveness of the proposed algorithms.The studies are expected to provide insights into the related research on multi-task optimization problems.