A Dissertation submitted to the FSU Dept.of Computer Science
and successfully defended on March 29, 1999.

R.S. Renner*
The Florida State University Department of Computer Science Tallahassee, FL

Ensemble networks have been receiving considerable attention within the last few years. Most existing models are created with linear networks. Ensembles of linear networks have demonstrated improved performance over individual networks, but linear models have limited capacity problems. Ensembles of more complex well-trained networks offer a promising alternative. Unfortunately, the computational expense involved in training large numbers of well-trained networks may be prohibitive. An ensemble of non-linear feed-forward neural networks generated by a constructive algorithm is presented. The ensemble method presented exhibits better generalization than linear ensembles, and shows promise toward a reduction in time-complexity over well-trained ensembles. The problems addressed in this research are: generalization of non-linear data, time-complexity, structural dilemmas, model creation, and model combination. The Neural Network Ensemble Simulator (NNES) is also introduced as a simulation tool for managing ensemble experiments. NNES provides routines for ensemble creation, selection, combination, and analysis.
* Present Address: Dept. of Computer Science
California State University - Chico
Chico, CA 95929-0410
E-mail: renner@ecst.csuchico.edu

Thanks to Lutz Prechelt for access to the PROBEN1 data repository,
and Scott Fahlman and Marc White for access to CMU's repository and CNNS.