Performance of Complex-Valued Multilayer Perceptrons Largely Depends on Learning Methods

摘要

Complex-valued multilayer perceptrons (C-MLPs) can naturally treat complex numbers, and therefore can work well for the processing of signals such as radio waves and sound waves, which are naturally expressed as complex numbers. The performance of C-MLPs can be measured by solution quality and processing time. We believe the performance seriously depends on which learning methods we employ since in the search space there exist many local minima and singular regions, which prevent learning methods from finding excellent solutions. Complex-valued backpropagation (C-BP) and complex-valued BFGS method (C-BFGS) are well-known for learning C-MLPs. Moreover, complex-valued singularity stairs following (C-SSF) has recently been proposed as a new learning method, which achieves successive learning by utilizing singular regions and guarantees monotonic decrease of training errors. Through experiments using five datasets, this paper evaluates how the performance of C-MLPs changes depending on learning methods.