RT and SLT provide a framework within which data analysis tools can be developed and compared. Both suggest not to focus on the minimization of an empirical error over existing data, since it is often ill-posed and may not lead to models with good predictive power. Minimizing a combination of empirical power and a penalty factor for solutions that are too complex is suggested as an effective alternative. Regularization networks and SVM are developed as particular cases. (21 refs.)
展开▼