Research On Differential Evolution Algorithms Using Surrogate Models And Reinforcement Learning Techniques

In a variety of optimization techniques,evolutionary algorithms have become common optimization methods in the engineering field by virtue of their advantages of self-organization,self-adaptation,self-learning and parallel search.However,when facing complex optimization problems such as high-dimensional,multi-modal and nondifferentiable,these existing algorithms have always some shortcomings such as easily falling into local optimum and time-consuming calculation.In addition,additional constraints may often cause that the optimization problem has a too narrow or even disjoint feasible region,which is very difficult to obtain sufficient feasible points.In developing optimization algorithms,one has to deal with the feasibility of points as well as optimality evaluated by objective functions.The above factors cause great computational challenges and resistance to the application of evolutionary algorithms.In the thesis,some efficient techniques and algorithmic framework,such as surrogate models,reinforcement learning and differential evolution,are adopted and developed to improve the performance of the optimization approaches.In the proposed algorithm,surrogate models are used to decrease the evaluation cost of fitness,whereas reinforcement learning technologies can obtain more heuristic information for highquality offspring.As a result,two novel differential evolution algorithms are developed for global and constrained optimization problems,respectively.Main works are presented as follows:Aiming at the large-scale global optimization problem,a differential evolution algorithm based on surrogate model and multi-strategy technology is designed.Firstly,in order to reduce the computational cost,the surrogate model with small computation cost is selected for fitness evaluation.In this method,both global and local surrogate models,instead of original ones,are constructed to guide population evolution.Secondly,in order to maintain the balance between the convergence and diversity of individuals,a multi-strategy method is used to generate offspring individuals.The evolution process is divided into two stages according to the average distance of individuals in the population,and differential mutation operators with different performance are selected according to their performance at different stages.Finally,simulation experiments are carried out on the benchmark functions,and the results show that the proposed algorithm has better performance and robustness than compared approaches.For constraint optimization problem,a differential evolution algorithm is designed by embedding reinforcement learning technology as well as a new constraint handling technology.Firstly,according to the characteristics of the constraint,a heuristic constraint handling technique is presented,which can effectively promote the evolution of infeasible individuals to feasible regions.Secondly,in order to balance the exploration and exploitation capacities of the algorithm,a reinforcement learning technology is introduced into the mutation operation.As a result,the mutation operators can be chosen self-adaptively.Finally,simulation experiments are carried out on the benchmark functions,and the results show that these hybrid techniques effectively improve the performance of the algorithm.