H. D. Navone, P. M. Granitto, P. F. Verdes, and H. A. Ceccatto, "A learning algorithm for neural network ensembles," Inteligencia Artificial, Revista Iberoamericana de Inteligencia Artificial, vol. 12, pp. 70--74, 2001.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....data in the training and testing phases the overall generalization capabilities of the network can be tuned. The purpose of this paper is to compare the generalization capabilities of the SP methodology on a standard benchmark problem. For these purposes Freidman s add10 [4] or Freidman#1 [5] 6][7]) problem was selected. The add10 function is defined as follows: f(x 1 , x 10 ) 10 sin(#x 1 x 2 ) 20(x 3 0.5) 10x 4 5x 5 # (1) where # is zero mean unit variance Gaussian noise. The inputs x 1 , x 10 are sampled independently from a uniform [0, 1] distribution. The ....

....1: ADD10 Benchmark Results Train Test SSE 5.4 SSE SP 50 350 2.5 2.1 SP 100 300 2.4 2.3 SP 200 200 2.7 2.0 SP 300 100 2.6 2.0 SP 350 50 2.9 1.9 GMDH 200 200 3.9 1.4 MARS 200 200 5.4 1.0 NBAG 200 200 4.5 1. 2 data points [8] In pervious benchmark studies on the add10 problem [5] 6] [7], the neural networks used (200,200,1000) for the training, testing and generalization data respectively. The sizes of these sets was varied between (50,350,1000) and (350,50,1000) as shown in table I. Table I tabulates the sum squared error results for the unseen or generalization data. For ....

H. D. Navone, P. M. Granitto, P. F. Verdes, and H. A. Ceccatto, "A learning algorithm for neural network ensembles," Inteligencia Artificial, Revista Iberoamericana de Inteligencia Artificial, vol. 12, pp. 70--74, 2001.

....so far minimize, in a global way, some particular error function. A different approach is to generate an ANN ensemble through the sequential aggregation of individual predictors, where the learning process of a new ensemble member is validated by the previous stage aggregate prediction performance [9,10]. That is, the early stopping method is applied by monitoring the generalization capability of the n stage aggregate predictor plus the n l network being currently trained on the validation data set V . In this way one retains the simplicity of independent network training and only the validation ....

....may have poor generalization capabilities. SECA is a stepwise optimization technique, and a known problem with these heuristics is that during the optimization process they cannot review the choices made in the past. An alternative is accepting only members that improve the ensemble performance [9], although this may lead to some overfitting. An intermediate solution is weighting the ensemble members instead of rejecting them if they do not improve the overall ensemble performance. This allows us to reduce the influence of bad choices made in the past by simply giving smaller weights to ....

H. Navone, P. Granitto, P. Verdes and H. Ceccatto, "A Learning Algorithm for Neural Network Ensembles", Revista Iberoamericana de Inteligencia Artificial 3, 70-74 (2001)

....This inclusion improved performance by incorporating dependencies on other variables not directly measured, although probably limiting the spatial validity of the developed models. In particular, to model the Sonic log we used a recently developed algorithm for building neural networks ensembles[3]. Here we approach the same problem using the Kernel Adaline algorithm[11] Both methods have shown excellent performance when tested on other real and artificial regression problems. We also develop linear regression models to be used as a basis for comparison. The work is organized as follows: ....

....a set of ANNs with both reasonably good (individual) generalization capabilities and distributed predictions for the test points. This problem has been considered in several recent works in the literature[6,10] In [17] we used an alternative way of generating an ANN ensemble, proposed in [3], which leads to ensemble members that are both accurate and diverse. The method essentially amounts to the sequential aggregation of individual predictors where, unlike in standard aggregation techniques which combine individually optimized ANNs[10] the learning process of a (potential) new ....

H.D. Navone, P.F. Verdes, P.M. Granitto and H.A. Ceccatto, "A Learning Algorithm for Neural Networks Ensembles", in G. P. Henning, ed., ASAI'2000.

....will improve performance by incorporating dependencies on other variables not directly measured, although probably limiting even more the spatial validity of the developed models. In particular, to model the Sonic log we use a recently developed methodology for building neural networks ensembles[3], which has shown excellent results when tested on other real and artificial regression problems. We also develop linear regression models to be used as a basis for comparison. The work is organized as follows: In Section 2 we describe the available data and the applied preprocessing. In Section ....

....generating a set of ANNs with both reasonably good (individual) generalization capabilities and distributed predictions for the test points. This problem has been considered in several recent works in the literature[6,10] We use here an alternative way of generating an ANN ensemble, presented in [3], which leads to ensemble members that are both accurate and diverse. The method essentially amounts to the sequential aggregation of individual predictors where, unlike in standard aggregation techniques which combine individually optimized ANNs[10] the learning process of a (potential) new ....

H.D. Navone, P.F. Verdes, P.M. Granitto and H.A. Ceccatto, "A Learning Algorithm for Neural Networks Ensembles", in G. P. Henning, ed., ASAI'

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC