Simple memetic computing structures for global optimization
Ilpo Poikolainen · Jyväskylä University Digital Archive (University of Jyväskylä) · 2014
During the recent years, Memetic Computing (MC) and Memetic Algorithms (MAs) have been drawing increasing attention in the scientific community. While MAs, by classic definition, include a population based algorithm and a local search, MC structures include several algorithmic components (memes) within a co-operative framework. Using several algorithmic components is preferable, over a single component, when wisely selected to complement each other. Along with the selection of these components, the designer needs to define a memetic structure, which determines how the components interact during the optimization process. Defining this structure is crucially important in order to achieve a robust algorithmic framework with good balance between exploration and exploitation. This thesis analyzes MC structures, focusing on the concept of simplicity, and understanding the role of each component. Several MC structures are presented in the included articles. First, a simple memetic structure is studied in depth, and its performance is enhanced by modified variants. This simple structure contains a resampling mechanism inspired by the exponential crossover of Differential Evolution (DE). The resulting algorithms retain the properties of the original implementation, in terms of simplicity and memory requirements, while remaining competitive against more complex state-of-art algorithms. Differential Evolution is studied in depth as some novel MC structures designed in this thesis are based on a DE logic. DE is a versatile optimization algorithm which can be applied to a wide range of problems. The overall simple structure is achieved by adding components and/or modifying operators from the original DE scheme. Finally, motivated by the philosophy that the role of each part of an algorithm should be clear to the designer and the algorithm should be tailored around the problem features, a novel DE based MC scheme is introduced. The MC scheme estimates the multimodality of an optimization problem, and detects the most interesting areas of the decision space in order to intelligently guide the initial population sampling for DE.