Cascade Correlation (Cascor) has proved to be a powerful method for training neural networks. Cascor, however, has been shown ont to generalise well on regression and some classification problems. A new Cascade network algorithm employing Progressive RPROP (Casper), is proposed. Casper, like Cascor, is a constructive learning algorithm which builds cascade networks. Instead of using weight freezing and a correlation measure to install new neurons, however, Casper uses a variation of RPROP to train the whole network. Casper is shown to produce more compact networks, which generalise better than Cassor.
展开▼