| Summary: | In the last ten years optimization in industrial applications has obtained an increasing attention. In particular it has been demonstrated useful and effective in the solution of control problems. The implementation of an optimization algorithm on a real-time control platform must cope with the lack of a full power computer, thus it must use a very low amount of memory and computational power. On the other hand the presence of nonlinearities, sensors and approximations injects in the signals of the control loop some noise, resulting in a noisy fitness function to be optimized. In this work both issues are addressed in order to show how a novel algorithmic design can arise from the solution of these implementation problems, often underestimated in the theoretical approach. This thesis proposes a set of novel algorithmic solutions for facing complex real-world problems in control engineering. Two algorithms addressing the optimization in the presence of noise are discussed. In addition, a novel adaptation system inspired by estimation of distribution paradigm is proposed to handle highly multimodal fitness landscapes. A crucially important contribution contained in this thesis is the definition of compact Differential Evolution for optimization problems in presence of limited hardware. Finally an evolution of the latter algorithm in the fashion of Memetic Computing is proposed with reference to an industrial application problem.
|
|---|