Preference-inspired coevolutionary algorithm with sparse autoencoder for many-objective optimization

doi:10.1007/s00500-023-09050-7

Return

Preference-inspired coevolutionary algorithm with sparse autoencoder for many-objective optimization

2023-08-17

PRE

DOI：10.1007/s00500-023-09050-7

Original

Original request for help

Save

Abstract

En 中文

The function of most multi-objective algorithms (MOEAs) is to provide an overall trade-off Pareto front to the decision makers (DMs). But DMs actually tend to have a preference for a specific subset of the Pareto front. So combing decision makers' preference information with many-objective optimization methods has recently been a hot survey topic in the research field of many-objective optimization. The preference-inspired coevolutionary algorithms using goal vectors (PICEA-g) coevolve a family of decision makers' preferences called preference set together with a population of candidate solutions. It is a classic and effective algorithm, but it has shortcomings in distribution and there is a waste of computing resources. So we use a sparse autoencoder as a controller to perform feature processing on the target solution set in the process of coordinated evolution and modify the fitness assignment formula to improve the diversity. The combination of PICEA-g and sparse autoencoder framework enhance the convergence and expand the ability of the PICEA-g algorithm. The preference-inspired coevolutionary algorithm with sparse autoencoder (PICEA-g/SAE) is evaluated by three widely used benchmark suites and compared with nine classic multi-objective evolutionary algorithms to prove the advantages of the sparse autoencoder framework. The experimental results show that PICEA-g/SAE could have a good performance on many DTLZ test problems. Moreover, PICEA-g/SAE on UF1-9 and WFG1-9 test problems in low dimensions could have good convergence, diversity, and spread compared with other algorithms.

Keywords:

Many-objective optimization

Preference-inspired coevolutionary algorithms

Sparse autoencoder

Decision makers