Relating binary and continuous problem entropy to backpropagation network architecture

摘要

The research presented focuses on characterization of problems, with emphasis on process control data, as they relate to backpropagation neural network architectures, particularly contrasting one and two hidden layers networks. The effects of architecture alterations on network convergence, performance, and stability are empirically examined for continuous and binary domains. Shannon entropy for binary sets and statistical complexity for continuous sets are calculated and related to optimal architecture.