Fuzzy Logic and Piecewise-Linear Regression
Abstrakt
McNaughton theorem of fuzzy logic states that the functions on the unit cube which admit representation by formulas of the Lukasiewicz propositional logic are exactly all continuous piecewise-linear functions (CPLFs) with integer coefficients. The paper discusses an approach based on the linear regression in this context. A CPLF and corresponding polyhedral partition of the input domain that best approximate given sample data are sought, considering the trade-off between the accuracy of the approximation and the complexity of the resulting formula. The proposed method is scalable and combines a kd-trie space-partitioning data structure, a linear regression and a clustering supported by hypothesis testing.