Extending and benchmarking the CasPer algorithm

Abstract

The CasPer algorithm is a constructive neural network algorithm. CasPer creates cascade network architectures in a similar manner to Cascade Correlation. CasPer, however, uses a modified form of the RPROP algorithm, termed Progressive RPROP, to train the whole network after the addition of each new hidden neuron. Previous work with CasPer has shown that it builds networks which generalise better than CasCor, often using less hidden neurons. This work adds two extensions to CasPer. First, an enhancement to the RPROP algorithm, SARPROP, is used to train newly installed hidden neurons. The second extension involves the use of a pool of hidden neurons, each trained using SARPROP, with the best performing selected for insertion into the network. These extensions are shown to result in CasPer producing more compact networks which often generalise better than those produced by the original CasPer algorithm.