Research On Improvement Of Differential Evolution Algorithm Based On The Tradeoff Between Exploration And Exploitation

Intelligence optimization has become one of the hottest topics in the research field ofevolutionary computation which solves optimization problems in heuristic manners.Compared with traditional optimizers like Newton’s method and conjugate gradient method,these heuristic algorithms have many virtues, such as independence on the gradient ofobjective functions, no requirement on the differentiability and continuity of objectivefunctions, as well as desirable global optimization ability. Due to these viruses, intelligentoptimizers have been widely applied to solve various complex optimization problemsencountered in scientific and engineering practice. Among these intelligent optimizers,differential evolution algorithm (DE) gains more focus for its simpleness, efficiency, andeasy implement. Now DE is mainly applied in the filed of continuous optimization, due toadvantages in very few control parameters, high reliability and strong robustness, it hasobtained great application in multi-objective optimization, constrained optimization anduncertain optimization problems. However, DE usually suffers from stagnation andpremature convergence when being used to solve complex optimization problems and multimodal optimization. This is a common problem of intelligent optimization algorithms.More studies are still focused on analyzing DE’s different components and putting forwardimprovement strategies from various perspectives such as control parameters, populationstructure and individual learning mechanisms.Obviously these methods can not work outfor the given problems.In view of the above several issues, this dissertation made an in-depth research on thedifferential evolution. Some new variants are given to improve the performance of DE byadjusting the control parameters dynamically. In contrast to Genetic Algorithm (GA) andParticle Swarm Optimization (PSO), the global convergence of DE deserves comprehensiveinvestigation. With respect to current research, how to balance the exploration andexploitation abilities of DE is a key point. The global convergence criteria of randomalgorithm with a finite population are given to prove that the conditional DE algorithm cannot guarantee global convergence. A new DE variant is proposed, which incorporates threemechanisms into the traditional DE algorithm. They are Gaussian mutation,diversity-triggered reverse sampling, and fast exploitation by a small DE population.Theoretical analysis and experimental results show that not only the global convergence can be guaranteed but also desirable optimization performance can be achieved via theproposed DE algorithm. The conclusion part summarizes the research in this paper andpoints out important and valuable research lines for future work regarding the design ofintelligent optimizers with the guidance of exploration-exploitation tradeoff.