Optimization Methods For Problems With Redundant Variables And Uncertain Objective Functions

This paper focuses on the multi-objective optimization problem with redundant variables and uncertain objective functions(MOPRVIF),and how to improve the computational accuracy of the objective function and the ability of the algorithm to perform deep intelligent search.In this regard,this paper proposes a dual data-driven multi-objective optimization method,which uses deep learning techniques to solve the accuracy problem and machine learning to solve the problem of intelligent search capability.The main work of this paper is as follows:(1)In the data pre-processing stage,redundant variables(ERV)are eliminated.The feature variables and the objective function are interconnected by similarity and correlation mining,aggregating different variables that share the same change trend and retaining variables with high importance,making the data information more condensed and facilitating subsequent problem solving.(2)Objective Function Construction(OFC).Using the types and topologies of deep learning models as building elements,the construction and optimization of objective function learning models(OFLM)are performed by artificial intelligence optimization algorithms.A variable-length encoding and decoding method and an evolutionary operation based on the idea of group optimization are designed.(3)Select the evolution operator(SEO).In response to the traditional optimization algorithm’s single evolutionary operation and lack of flexibility in deep search,an evolutionary operator adaptive optimizer(EOAO)is designed based on a machine learning model to improve the deep intelligent search capability of multi-objective evolutionary algorithms.(4)A standardized optimization architecture is formed based on a data-driven approach.MOEA is the main line of the architecture,ERV is the data preparation.When constructing the objective function model,high accuracy OFLM is an important guarantee for reliable solution results.The ability of EOAO to adaptively adjust the evolutionary operator during search is an important guarantee for deep intelligent search.MOEA combines OFLM and EOAO to form the final solution algorithm-Dual Data Driven Multi-Objective Evolutionary Algorithm(DDMOEA).In this paper,validation cases and application cases are selected for simulation experiments of data-driven optimization.The validation case is a drug candidate selection to determine the best values or intervals of molecular descriptor variables that would allow the candidate compounds to have better efficacy targets.The application case is a wild blueberry cultivation problem,where the optimal environmental parameters for cultivation are explored to maximize crop yield and minimize human intervention.Both experimental cases are similar with redundant variables and uncertain objective functions,but different except that the domain is different and the validation case is more complex.The two experimental cases demonstrate that the data-driven design is successfully applied to two different types of MOEAs,and the combined effect of convergence and diversity in solving the problem has certain advantages,and the proposed data-driven optimization method has certain effectiveness,practicality and extensibility.