A Tuning Free Differential Evolution Meta Framework for Unconstrained Optimization Problems

Abstract

Evolutionary Algorithms (EAs) are population-based algorithms, they use an iterative evolutionary process to explore the solution space and reach the optimal solution. Differential Evolution (DE) is added to the group of EAs in 1995. DE uses differential mutation and crossover operations to generate new possible solutions, which are controlled by the parameters mutation scale factor (F) and crossover rate (CR). There are different types of mutation and crossover strategies, for DE, proposed by various researchers and these lead to variants of the original DE algorithm. Another parameter that plays a major role in DE solving an optimization problem is the Population Size (NP), the number of possible solutions that are maintained at any given point of time. In order to use DE as a tool to solve any real-world optimization problem, its parameters need to be set with appropriate values. The best values of these parameters keep on changing with respect to the problem under consideration. Even though some initial guidelines for setting the parameter values are provided by the researchers, there is no single value that will work for all problems. It will be beneficial if the algorithm can automatically decide/set the appropriate values for the parameters depending on the problem being solved. This led to the research direction of parameter control, which started around 2004 for DE algorithm. Most of the parameter control mechanisms available in DE literature concentrates on F and CR. In order to set the appropriate value for a control parameter, its effect on the performance of the algorithm needs to be studied. The aim of this research work is to design and implement a tuning free DE algorithm, which can control its major subset of parameters. The first step is to study the various adaptation strategies used in the existing literature. The next two objectives of the research were understanding the impact of F and CR parameters on the nature of convergence of the DE algorithms and on proposing relevant adaptation ...