Aprendizado supervisionado usando redes neurais construtivas

Abstract

Constructive neural learning is a neural learning model that does not assume a fixed network topology before training begins. The main characteristic of this learning model is the dynamic construction of the network s hidden layers that occurs simultaneously with training. This work investigates three topics related to constructive neural learning namely algorithms for training an individual TLU, constructive neural algorithms for two class problems and constructive neural algorithms for multiclass problems. The first research topic is approached by discussing a few TLU training algorithms, namely Perceptron, Pocket, Thermal, Modified Thermal, MinOver and BCP. This work approaches constructive neural learning for two class classification tasks by initially reviewing Tower, Pyramid, Tiling and Upstart algorithms, aiming at their multiclass versions. Next five constructive neural algorithms namely Shift, Offset, PTI, Perceptron Cascade and Sequential are investigated and two hybrid algorithms are proposed: Hybrid Tiling, that does not restrict the TLU s training to only one algorithm and the OffTiling, a collaborative approach based on Tiling and Offset. Multiclass constructive neural learning was approached by investigating TLUs training algorithms that deal with multiclass as well as by investigating multiclass versions of Tower, Pyramid, Tiling, Upstart and Perceptron Cascade. This research work also describes an empirical evaluation of all the investigated algorithms conducted using several knowledge domains. Results are discussed and analyzed.