Genetic Programming in Application to Flight Control System Design Optimisation

Abstract

1.1 Evolutionary algorithms EAs are often considered as an example of artificial intelligence and a soft computing approach. Their unique ability to search for complete and global solutions to a given problem makes EAs a powerful problem solving tool which combine such important characteristics as robustness, versatility and simplicity. Historically, there exist several branches of EAs, namely Genetic Algorithms, Genetic Programming, Evolutionary Programming and Evolutionary Strategies. Their development started independently in the 1960s and 70s. Nevertheless, all of them are based on the same fundamental principle evolution. ‘Evolution’ is used here in its Darwinian sense, the advance through ‘survival of the fittest’. Despite of a remarkable simplicity, EAs have proven to be capable of solving many practical tasks. The first and obvious application is numerical optimisation (minimisation or maximisation) of a given function. However, EAs are capable of much more than function optimisation or estimation of a series of unknown parameters within a given model of a physical system. Due to, in a large part, their stochastic nature, EAs can create such complex structures as computer programs, architectural designs and neural networks. Several applications of EAs have been known to produce a patentable invention (Koza et al., 1996, Koza et al., 1999 and Koza et al., 2003). Unfortunately, such a truly intelligent application of EAs is rarely used for practical purposes. GP and similar algorithms often require a supercomputing power to produce an optimal solution for a practical task. This may be overcome, at least partially, by narrowing the search space. A general engineering design practice is to propose a new design based on existing knowledge of various techniques (not uncommonly even from other fields) and no less important, intuition. Following this, the proposal is analysed, tried on a real system or its mathematical model, findings and errors are analysed again, the design is modified (or rejected) and the process continues until a satisfactory solution is found. EAs work basically on the same principle, although, obviously, using less analytical analysis but more trial-and-error approach. It was found, however, that the process of selecting the most suitable solutions at each stage and producing the next iteration variants is, overall, largely intelligent and heuristic. EAs are capable to answer not only the question how to do something (how to control, in particular), but also the question what to do in order to meet