Research On Transfer Enhancement For Multi-objective And Multi-task Optimization Algorithm

Most optimization problems in the fields of science,technology and engineering practice are optimization problems,some of which have a connection,called multi-task optimization problems.They are usually solved by evolutionary algorithm based on population iterative optimization.However,solving each optimization problem one by one would be a waste of time and resources.Multi-task optimization algorithm is an important method to solve this kind of problems.It can carry out appropriate knowledge transfer to the current optimization problem based on historical experience knowledge,and then promote the search for the optimal solution set of the problem to be optimized.At present,the difficulty is how to reduce the negative transfer phenomenon to improve the optimization efficiency of multi-task evolutionary optimization.Therefore,this paper proposes two novel transfer enhancement strategies to ensure that the optimal solution sets of multiple optimization problems can be obtained simultaneously under the limited number of true function evaluations.1.Multi-objective multi-task optimization algorithm with kernel technique.In order to reduce the occurrence of negative transfer in the process of multitask evolution,a kernelized autoencoder is suggested to capture the relationship between data so that tasks can co-evolve.By mapping different tasks to the reproducing kernel Hilbert space,the nonlinear relationship between the highdimensional space and the mapping space is constructed,and the effective knowledge transfer between tasks is realized.In addition,the elite strategy is also used to update the population,which further reduces the negative transfer probability.In the experiment,it tests on nine multi-objective multitask benchmark test functions and the experimental results show that the proposed algorithm has better optimization ability than other algorithms for solving multitask optimization problems under the same experimental conditions.2.Multi-objective and multitask optimization algorithm based on two-stage strategy.In the first stage,the individuals with a large degree of correlation between the excellent individuals of the target task were transferred through the grey relational analysis technology.And the differential evolution algorithm was used to do a small number of iterations to reduce negative transfer to the target task.In the second stage,the archive sub-population with larger correlation degree in the first stage is combined with the nonlinear relationships constructed by the nucleated autoencoder.Each generation of parents individuals is updated using the elite strategy.Compared with a number of popular algorithms,the proposed algorithm performs better in nine multi-objective multitask benchmark functions.