Extensiones del enfoque genético iterativo para el aprendizaje de reglas difusas

• Autores: David García Muñoz
• Directores de la Tesis: Raúl Pérez Rodríguez (dir. tes.) , Antonio González Muñoz (dir. tes.)
• Lectura: En la Universidad de Granada ( España ) en 2015
• Idioma: español
• Tribunal Calificador de la Tesis: Francisco Herrera Triguero (presid.) , Jesús Alcalá Fernández (secret.) , Dimitrios Stavrakoudis (voc.) , Luis Magdalena Layos (voc.) , María Jesús de la Fuente Aparicio (voc.)
• Enlaces
  • Tesis en acceso abierto en: DIGIBUG
• Resumen

  • This dissertation makes sense in the framework in which it is necessary to handle large amounts of data, some of them including vague or imprecise information.

    Particularly, the methods here proposed are based on the Iterative Rule Learning (IRL) approach, characterized by the use of a sequential covering strategy [1] together with a genetic algorithm in order to learn fuzzy rules.

    Considering this fact, we focus our efforts in two main objectives: on the one hand to consider the indirect relevance of the input attributes in the learning process and on the other hand to improve the IRL model used by a basic fuzzy rule-based learning algorithm (NSLV) to be able to iteratively review the learned knowledge.

    In this sense, the first objective is achieved through the use of feature construction [2, 3] techniques which allows the extraction of additional information resulting from the combination between the original variables. So, the learning algorithm handles not only the information given by the input variables, but also that given by the new constructed features. Following this idea, in this work we have presented three methods including feature construction techniques: one of them using relations in the antecedent of fuzzy rules (NSLV-R), another one using functions in the antecedent of such rules (NSLV-F) and the last one combining both relations and functions (NSLV-FR).

    It was experimentally proved that feature construction techniques work well when looking for accurate models, with an interpretability level nearer to natural language. Nevertheless, some well-known interpretability measures [4] refers to the number of rules of rule bases and the number of conditions per rule (also per rule base), as key elements in order to consider a ruleset as interpretable.

    It is inside this searching process of interpretable rule bases where the second main objective takes place. The ability of a fuzzy rule-based learning algorithm to review the knowledge as part of its own learning process, allows tuning the knowledge in each step. In this way, our proposal including knowledge review (NSLV-AR), decides in each iteration whether to replace or not a previously learned set of rules.

    Finally, the last proposal (SLAVE3), arises from the need to integrate the ideas previously exposed into a new model achieving a good trade-off between accuracy and interpretability. So, on the one hand we were looking for an algorithm with a high level of accuracy, similar to those using feature construction, and on the other hand, which also were able to improve the interpretability (by reducing the complexity at a rule/rule base level), when compared with those last ones.

    All the algorithms previously mentioned demonstrate their performance through an exhaustive experimental study.