Research And Application Of Supervised-surrogate Assisted Evolutionary Algorithms

As an important branch of Artificial Intelligence,Evolutionary Algorithms(EAs)have become a main way to solve complex optimization problems.The individual fitness evaluation is an important step of EAs and its efficiency and accuracy affect the performance of EAs.The main difference between Surrogate Assisted Evolutionary Algorithms(SAEAs)and traditional EAs is the difference of the individual fitness evaluation method.Traditional EAs use exact function for fitness evaluation,while SAEAs use the surrogate model.Generally,the time complexity of fitness evaluation using surrogate model is far less than that using the exact function.Therefore,compared with traditional EAs,SAEAs have a higher algorithm efficiency and are often used to solve complex optimization problems that require high algorithm efficiency.However,the study of SAEAs is still in infancy and much more effort should be put to improve the solution efficiency and solution quality of SAEAs.This thesis mainly includes the following contents.(1)In this thesis,a framework of Supervised-Surrogate Assisted Evolutionary Algorithms(S-SAEAs)is proposed.The algorithms take the evaluation accuracy as a supervision signal.Based on this supervision signal,a new model management strategy and a new individual generation strategy are designed.The former improves the solution efficiency by adjusting the update frequency of the surrogate model,while the latter improves the solution quality by maintaining the diversity of the population.(2)To further verify the performance of S-SAEAs,the thesis applys S-SAEAs to solve the practical problem of local dimming.Based on the framework of S-SAEAs,a Supervised-Surrogate Assisted Particle Swarm Optimization-Local Dimming(SSAPSO-LD)is designed.Experiments show that S-SAPSO-LD can improve the algorithm efficiency while ensuring the output image quality.