Komunitas Soft Computing Indonesia

Colloquium disajikan oleh Pitoyo Hartono, Future University-Hakodate, Hakodate City, Japan

Resume

Recently, several models of neural networks ensemble have been proposed, generally with the objective of achieving a higher generalization performance compared to the singular neural network. Some of the ensembles, represented by Boosting and Mixture of Experts, proposed mechanisms to divide the learning space into a number of sub-spaces and assign each sub-space into one of the ensemble's member, hence the learningload of each member can be relaxed, leading to better convergence and performance. In this paper, an ensemble of linear perceptron (ELP) with an algorithm that effectively divides the learning space in a linear manner, and assign the classification task of each sub-space to a linear perceptron, is proposed. The objective of this algorithm is to achieve a linear decomposition of a nonlinear problem through an automatic divide and conquer approach. In ELP, in addition to the ordinary output neurons, each linear perceptron has an additional neuron in its output layer. The additional neuron is called confident neuron, and produces an output that indicates the gconfidence levelh of the perceptron with regard to its ordinary output. An output of the perceptron which has a high confidence level can be considered as a reliable output, while an output with low confidence level is an unreliable one. The proposed ELP is equipped with a competitive mechanism for learning space division based on the confidence levels and at the same time to train each member to perform in the assigned sub learning space. The linearity of each member also enables us to analyze the division of the problem space that can be useful in understanding the structure of the problems and also to analyze the overall performance of the ensemble.