| This research focuses on the classification of biomedical data using an optimized extension to the fuzzy adaptive logic network (FALN). The FALN exploits the advantages of geometric representations of data (geometry-based processing) and fuzzy logic-oriented operations (logic-based processing) on those representations. The construction of an optimized FALN (OFALN) involves the selection of a heterogeneous structure (collection) of learning architectures (for example, multilayer perceptrons (MLP) and radial basis functions (RBF)) for the geometry-based processing subsystem.;While error backpropagation is adequate for optimizing an individual MLP, since the error may be described using a differentiable function, it cannot be used for the learning of an optimal heterogeneous structure. One possible solution to this structure optimization problem is to use a genetic algorithm (GA) to find an optimal combination of MLPs and RBFs. My thesis describes the FALN, MLP, RBF, and GA and also describes the structure optimization strategy used for OFALN. Finally, the use of OFALN, which is optimal with respect to the model implemented using the GA methodology, to classify a biomedical dataset is presented including a discussion of the performance results. |
