EVOLUTIONARY DESIGN OF GEOMETRIC-BASED FUZZY SYSTEMS

The main contribution of this thesis is the definition of a new model of fuzzy system where the exponential growth of the number of rules with respect to the number of input variables is reduced, with an efficient representation for the design, using evolutionary algorithms. In the proposed model, the partition of the input space is not defined as a regular structure built as the intersection of the linguistic labels of input variables, as usual in fuzzy systems, but in terms of multidimensional regions, each one associated with a single fuzzy rule.The partition is defined based on well known concepts of computational geometry: the Voronoi diagrams and the Delaunay triangulations. The fuzzy system defined in terms of this partition has a clear and appealing structure. The representation of the individuals for evolutionary algorithms is simple, since each region in the multidimensional input space is represented with a single point. This geometric representation allows the use of geometric based operators for evolution. As an added advantage, the model allows an interesting approach for the inclusion of a priori knowledge about the solution of the problem in the individuals before and during the evolution.Experimental results on evolutionary design of Voronoi based fuzzy systems are presented in two control problems: an inverted cart pole system and a typical robot control application. The approach is extended to the design of recurrent Voronoi-based fuzzy systems. This extension is evaluated in two other control problems: a system identification problem, where the outputs are defined in terms of past inputs and outputs, and a problem from evolutionary robotics, where the ability to introduce a priori knowledge in the form of recursive rules is demonstrated