Research And Design Of Evolutionary Dynamic Multi-objective Optimization Algorithm Based On Knowledge Transfer

Multi-objective optimization problem is one of the most important and common problems in engineering application and scientific research.It widely exists in many practical scenarios such as control system,investment portfolio,production allocation and commodity pricing.Traditional mathematical programming methods such as the weighted method,the constraint method,and the ideal point method,etc.,can only obtain a single optimal solution under a given weight,in dealing with the complex problems containing multiple objectives.In addition,these classical programming methods highly rely on the weight and the order of objectives,they are thus impractical in solving multiobjective optimization problems.In contrast to traditional programming methods,evolutionary algorithms that are inspired by the principles of biological evolution have been successfully applied to solve multi-objective problems in the past few decades.Existing popular multi-objective evolutionary algorithms such as NSGA-II,SPEA,MOEA/D and IBEA have been adopted in a large number of researches and applications,and their effectiveness has been verified on many multi-objective benchmark and realworld problems.However,when solving a multi-objective optimization problem in reality,the objective function,the constraint function and the related parameters are very likely to change over time.For example,the optimization of resource allocation in power system,the optimization of investment scheme in stock trading,and the dynamic path planning in vehicle transportation,are all facing the dynamic change of problem properties.Therefore,in reality,the study of dynamic multi-objective optimization problem(DMOP)is a non-trivial task.Since DMOP has the characteristic of constantly changing with time or environment,the evolutionary algorithm not only needs to obtain the Pareto-optimal set(POS)within the given computational budget,but also needs to respond to dynamic changes as soon as possible,and search for the POS of current environment before the next change arrives,which greatly increases the difficulty of solving the problem.Due to the discussions above,we can see that the existing multiobjective optimization algorithms cannot find the moving POS in dynamic environment.Therefore,in order to solve DMOP,additional designs are needed to enable the evolutionary algorithm to adapt to the changes of the problem in the optimization process.Based on comprehensive investigation in the direction of dynamic multi-objective optimization,this thesis considers the optimization experience of solving historical problems,and focuses on the core problem of how to predict new environmental changes through knowledge transfer,then explores how to solve dynamic multi-objective optimization problems efficiently and effectively.Specifically,main contributions of this thesis can be summarized as the following three aspects:(1)Firstly,for solving DMOP with predictable change patterns,considering existing prediction algorithms cannot make full use of historical search information to improve prediction accuracy,a dynamic multi-objective optimization algorithm based on prediction is proposed to realize knowledge transfer before and after problem change in dynamic environment through autoencoding evolutionary search paradigm.The main contributions of this study are as follows: Considering the non-dominated solution set obtained before the change,then an efficient denoising autoencoder is derived to predict the overall moving directions of POS,which can quickly track the dynamic change of the problem and effectively maintain the diversity of the population in the evolution process;The proposed algorithm is compared with 3 different prediction methods to solve 10 DMOP test problems.The experimental results confirm the efficacy of the proposed method in achieving more accurate and diverse optimization performance in contrast to existing dynamic multi-objective optimization algorithms.(2)Further,based on the existing prediction based dynamic multi-objective optimization algorithms,for the key difficulty brought by different types of dynamics in different problem spaces,a general dynamic multi-objective optimization algorithmic framework based on multi-view prediction is proposed.The main contributions of this study are as follows: In contrast to the existing prediction algorithms which mostly track POS movement in the decision space,the proposed framework can simultaneously predict different change patterns in both decision space and objective space,so as to grasp all the impacts of environmental changes on problem solving as much as possible;Further,in order to realize the learning and prediction of different types of dynamic changes,a kernelized autoencoding model is derived in the Reproducing Kernel Hilbert Space to perform the proposed multi-view prediction,which is based on the knowledge transfer implemented by the autoencoding evolutionary search paradigm in the previous study;The proposed algorithmic framework is compared with 6 different prediction methods to solve 14 DMOP test problems.The experimental results show that the proposed multiview prediction can not only improve the optimization performance,but also ensure the diversity and robustness of the algorithm for solving a wide range of DMOPs.(3)Finally,based on the research and design of the above two parts of evolutionary dynamic multi-objective optimization algorithm,aiming at the application of dynamic multi-objective optimization algorithm in solving real-world DMOPs,a dynamic multiobjective recommender system modeling method is proposed.Meanwhile,a dynamic multi-objective optimization algorithm for sequential recommendation is designed according to the real application scenarios.The main contributions of this study are as follows: A novel and flexible dynamic multi-objective modeling method is proposed in the context of sequential recommender system;an end-to-end recommendation algorithm,DMORec,is proposed to predict users’ dynamic preferences on different objectives.In addition,the proposed algorithm is compared with 6 recommendation algorithms of different categories on 4 real datasets of different magnitudes and different application scenarios.The experimental results show that DMORec can quickly respond to environmental changes and generate recommendation results in line with users’ dynamic preferences.