New Strategies for Initialization and Training of Radial Basis Function Neural Networks

In this paper we proposed two new strategies for initialization and training of Radial Basis Function (RBF) Neural Network. The first approach takes into consideration the "error" between the input vector p of the network and the x-axis, which are the centers of radial functions. The second approach takes into account the "error" between the input vector p and the network output. In order to check the performances of these strategies, we used Brazilian financial market data for the RBF networks training, specifically the adjusted prices of the 10 greater weighted shares in the Bovespa index at the time of data collection ‑ from April 8th, 2009 to October 31th, 2014. The first approach presented a 52% of improvement in the mean squared error (MSE) compared to the standard RBF network, while the improvement for the second approach was 38%. The strategies proved to be consistent for the time series tested, in addition to having a low computational cost. It is proposed that these strategies be improved by testing them with the Levenberg-Marquardt algorithm.