par Medina-Chico, Victor;Suárez, Alberto;Lutsko, James 
Référence Lecture notes in computer science, 2167, page (348-359)
Publication Publié, 2001
Référence Lecture notes in computer science, 2167, page (348-359)
Publication Publié, 2001
Article révisé par les pairs
| Résumé : | A global optimization algorithm is designed to find the parameters of a CART regression tree extended with linear predictors at its leaves. In order to render the optimization mathematically feasible, the internal decisions of the CART tree are made continuous. This is accomplished by the replacement of the crisp decisions at the internal nodes of the tree with soft ones. The algorithm then adjusts the param- eters of the tree in a manner similar to the backpropagation algorithm in multilayer perceptrons. With this procedure it is possible to generate regression trees optimized with a global cost function, which give a continuous representation of the unknown function, and whose architecture is automatically fixed by the data. The integration in one decision system of complementary features of symbolic and connectionist methods leads to improvements in prediction efficiency in both synthetic and real-world regression problems. |
