Effect of multi-hidden-layer structure on performance of BP neural network: Probe

As a multi-layer forwarding network, the back propagation neural network (BPNN) with manifold derived structures has been most widely used in artificial intelligence applications. Based on the given non-linear system and the BPNNs of varying internal structures, this paper quantitatively reports the findings in the correlation between the number of hidden layers and the BPNN performance. The selection of learning rate is also investigated using the 3-layer BPNN and the same non-linear system. Through the simulation results in this probe it finds that the BPNN performance is not improved notably or even degraded with the increase of hidden layers, and 3-layer (or 1-1-1) BPNN is identified as the best performer.