Uncertainties Formulated as a Classification Problem Applied to Chaotic System
Pettras L. B. dos Santos, Sandra Aparecida Sandri, Haroldo F. de Campos Velho · Mecánica Computacional (Asociación Argentina de Mecánica Computacional) · 2011
Uncertainty is one main concern on operational prediction systems, because the initial conditions are not precisely determine, model parameters are estimated with few information, and the phenomena itself is not completely formulated – for instance: in the numerical weather prediction, turbulence is not fully understood and/or described. A quantitative evaluation how good is the prediction can be called predictability. The ”bred vector” methodology can be applied to characterized classes of dynamics. Two neuro-fuzzy systems are employed as class dynamics classifiers: (a) ANFIS (Adaptive-Network-based Fuzzy Inference System) based on Takagi-Sugeno’s approach, (b) GUAJE (Generating Understandable and Accurate fuzzy models in a Java Environment) based on Mamdami’s scheme. The technique is applied to a chaotic system: three coupled waves in solar physics. A better classification performance is obtained using the ANFIS, but the automatic rules generated by the GAUJE are more easily interpretable.