Jen-Lun Yuan, Terrence Fine
We are developing a forecaster for daily extremes of demand for electric power encountered in the service area of a large midwest(cid:173) ern utility and using this application as a testbed for approaches to input dimension reduction and decomposition of network train(cid:173) ing. Projection pursuit regression representations and the ability of algorithms like SIR to quickly find reasonable weighting vectors enable us to confront the vexing architecture selection problem by reducing high-dimensional gradient searchs to fitting single-input single-output (SISO) subnets. We introduce dimension reduction algorithms, to select features or relevant subsets of a set of many variables, based on minimizing an index of level-set dispersions (closely related to a projection index and to SIR), and combine them with backfitting to implement a neural network version of projection pursuit. The performance achieved by our approach, when trained on 1989, 1990 data and tested on 1991 data, is com(cid:173) parable to that achieved in our earlier study of backpropagation trained networks.