The abstraction of real-life situations based on the inference of observed phenomena, usually by way of statistical methods, is the main goal of several scientific fields. More than a few statistical methods exist today; however, nonlinear regression and multigene algorithms are the more widely accepted ones. This paper proposes the use of a hybrid symbolic regression (SR) algorithm by combining concepts from evolutionary computing and nonlinear regression using kernel methods. This algorithm creates models that satisfactorily fit experimental data. Notably, experimental data are fundamental for the successful execution of inference and estimation tasks. This paper considers models whose inference corresponds to a nonlinear combination of a certain number of mathematical expressions. The performance of the proposed algorithm in the estimation of the values of variables is evaluated using six databases available in virtual repositories. Models generated by this algorithm show a significant reduction in the estimation error when compared to two other generated models: (1) the one obtained by running the original symbolic regression algorithm and (2) that which considers a linear relationship between the mentioned mathematical expressions.

genetic programming, inference, kernel methods, linear and nonlinear regression, symbolic regression.