Optimization of flight control parameters of an aircraft using genetic algorithms

Genetic Algorithms (GAs) are stochastic search techniques that mimic evolutionary processes in nature such as natural selection and natural genetics. They have shown to be very useful for applications in optimization, engineering and learning, among other fields. In control engineering, GAs have been applied mainly in problems involving functions difficult to characterize mathematically or known to present difficulties to more conventional numerical optimizers, as well as problems involving non-numeric and mixed-type variables. In addition, they exhibit a large degree of parallelism, making it possible to effectively exploit the computing power made available through parallel processing. Despite active research for more than three decades, and success in solving difficult problems, GAs are still not considered as an essential global optimization method for some practical engineering problems. While testing GAs by using mathematical functions has a great theoretical value, especially to understand GAs behavior, these tests do not operate under the same factors as real life problems do. Among those factors it is worth to mention two possible situations or scenarios: one is when a problem must be solved quickly on not too many instances and there are not enough resources (time, money, and/or knowledge); and the other is when the objective function is not "known" and one can only "sample" it. The first scenario is realistic in engineering design problems where GAs must have a relatively short execution time in achieving a global optimum and a high enough effectiveness (closeness to the true global optimum) to avoid several iterations of the algorithm. The second scenario is also true in the design of technical systems that generally require extensive simulations and where input-output behavior cannot be explicitly computed, in which case sampling becomes necessary. F1ight control design presents these two types of scenarios and, during the last ten years, such problems as structure-specified Hoo controllers design, dynamic output feedback with eigenstructure assignment, gain scheduled controllers design, command augmentation system design, and other applications have been targeted using genetic methods. Although this research produced very interesting results, none so far has focused on reducing the execution time and increasing the effectiveness of GAs. The efficiency and effectiveness of Genetic Algorithms are highly determined by the degree of exploitation and exploration throughout the execution. Several strategies have been developed for controlling the exploitation/exploration relationship for avoiding the premature convergence problem. While significant body of expertise and knowledge have been produced through several years of empirical studies, no research has reported the use of Bayes Network (BN) for adapting the control parameters of GAs in order to induce a suitable exploitation/exploration value. The present dissertation fills the gap by proposing a model based on Bayes Network for controlling the adaptation of the probability of crossover and the probability of mutation of a Real-Coded GA. The advantage of BNs is that knowledge, Iike summaries of factual or empirical information, obtained from an expert or even by learning, are interpreted as conditional probability expressions. It is important to highlight that our interest, which is motivated by the requirements of real applications, is the behavior of GAs within reasonable time bound and not the limit behavior. Related genetic algorithm issues, such as the ability to maintain diverse solutions along the optimization process, are also considered in conjunction with new mutation and selection operators. The application of the new approach to eight different realistic cases along the flight control envelope of a commercial aircraft, and to several mathematical test functions demonstrates the effectiveness of GAs in solving flight-control design problems in a single run