Discrete Particle Swarm Optimization for the Selection of Input Variables of Feed Forward Neural Networks

The selection of input variables to an artificial neural network is typically a manual, trial-and-error endeavor. Given that the optimization of the weights of any single neural network is generally very time-consuming, the process of variable selection can be quite tedious.  This work describes an approach for the automatic selection of relevant input variables. It uses discrete particle swarm optimization to select variables as well as other neural network structural elements (number of hidden layers and nodes and input lag structure) in order to optimize the topology of non-linear, multivariate time series networks.  The application of the technique will be illustrated with a case study using data available in the public domain.