Parallel genetic algorithms: an efficient model and applications in control systems

Muhammad, Ali. (1997). Parallel genetic algorithms: an efficient model and applications in control systems. PhD thesis, Southampton Solent University [Validated by Nottingham Trent University].

Abstract

Optimisation is an adaptiver process and it is widely applied in science and engineering from scheduling a manufacturing process to control of a spacecraft. Genetic algorithms are search and optimisation methods based on the mechanics of natural evolution. they provide promising results in non-linear and complex search problems. They have been proven to be parallel and global in nature. Genetic algorithms run slowly on sequential machines which is their mmmajor drawback. Most of the applications of genetic algorithm in engineering are in the area of design and schedule optimisation, where usually enough time is avialable to simulate the algorithm. The computer architecture is a main bottleneck since the sequential computation does not reflect the true spatial structure of the algorithm. There are a couple of parallel models and implementations available which improve the performance of these algorithms. The aim of this research is to develop a new model and/or to improve existing parallel models for real-time application of these methods in system identification and intelligent control. The desired features of this new model are: it should be independent of the optimisation problem, so it could be able to cope with a black box problem, it could be used in real-time applications, where the exact model of the system is unknown, and it should be implementable within the current technological framework. An extensive study of the current literasture on genetic algorithms has been carried out. A detailed review of the underlying theory of genetic algorithms has also been presented. A parallel model of genetic algorithm has been proposed and implemented on a transputer based system using ANSI C toolset for transputers. It has been tested for different strategies on the traditional suite of optimisation problems, ie DeJong's function and Deceptive functions. The results are compared using the performance measures proposed by DeJong. The performance and efficiency measures of the algorithm have been defined and worked out for the simulation results. The research has advanced the understanding of genetic algorithms as stochastic processes. A Markov chain based mathematical model has been developed. An informal study of convergence properties of the algorithm are presented from different points of view, ie time series, real analysis, Markov chains and metric topology. Gradient like information has been integrated into the genetic search in order to improve the performance and efficiency of the algorithm. A novel directional search method has been developed tested on the same set of problems and comapred using the same performance and efficiency measures as those reported in the recent publications. Unlike neural networks and fuzzy systems, genetic algorithms do not provide any general logic for system modelliung. Therefore system identification is achieved by using the fuzzy network for general logic and a genetic algorithm for parameter estimation giving as a result an evolving fuzzy network. This novel method has been applied to modelling of chaotic time series and it has been used to control a highly non-linear system, ie inverse pendulum. It is expected that with the advance of re-configurable electronics, evolutionary chips will be realised in the near future. They will play an important role in the development of genetic algorithms based control systems

Actions (login required)

View Item