Automatisation du processus de construction des structures de données floues

This thesis presents the automatic generation of fuzzy data structures. A fuzzy data structure is a fuzzy knowledge base without its inference engine. The generic name of “fuzzy knowledge bases” will be used throughout this thesis. The optimization tool used for the automatic learning is a genetically based algorithm. The first objective of this research is to prove the feasibility of the automatic generation of fuzzy knowledge bases (without the need of a human expert).

The genetic algorithms developed in this thesis follow two coding paradigms: a traditional binary coding and a new real/binary-like coding (hybrid coding). The evolution operators are adapted to each algorithm. In the hybrid coding the reproduction mechanism is made of two distinct parts: a specialized crossovers suited for the factual base (real coded part) and a traditional single point crossover used for the rule base (binary-like coded part). A comparative study on the learning performances of both algorithms, using synthetic data obtained from theoretical 3D surfaces, is done taking into account several performance criteria such as: the precision, the simplicity and the learning time of the genetically-generated fuzzy knowledge bases. From this comparative study, the hybrid coding emerged as the most efficient for most of the performance criteria, which prove the advantage of adapting a genetic algorithm to the optimization problem under study. This part resulted into the first publication presented in this thesis. One of the most tedious problems encountered in automatic learning using meta-heuristic algorithms (genetic algorithms being a part of the meta-heuristic algorithms family) is the premature convergence. In order to overcome these problems (only the hybrid approach is considered) several methods to improve the diversity within the population of solutions are developed. These methods use multiple reproduction strategies (using different crossover mechanisms) along with a crowded family strategy. These approaches showed their superiority when compared with the conventional reproduction strategies (using a single crossover mechanism through the entire evolution). Furthermore, a study on enhancing the performance of the genetic learning by improving the balance between exploration and exploitation within the individuals is done. This study showed the existence of evolution stages in genetic algorithms and also the influence of the exploitation/exploration levels on genetic learning performance. Exploration at the early stages of the evolution, followed by relaxed exploitation during the evolution stage and exploitation in the last stages is the order that improves the learning performance. Genetic learning on experimental data obtained from a tool wear monitoring application gave very satisfactory results. This part resulted into two more publications. An application of the evolutionary algorithms to the thermomechanical pulp and paper process (TMP) was performed, where the quality of the pulp is defined by the ISO brightness. The learning of the fuzzy knowledge bases is performed using input variables obtained from a Chip Management System (CMS ^©). The CMS ^© characterizes the quality of wood chips upfront of the TMP process using sensors such as: an RGB camera and near-infrared sensor. This approach allows an online prediction of the pulp quality, since no laboratory measurements are needed for the prediction. This part resulted into a fourth and fifth publication. Finally, a general discussion followed by a set of recommendations and conclusions close this thesis.