A Comparison Between ANN and RBFNN for Signal Demodulation in Photon Counting based Optical Wireless Communications

The use of silicon photomultiplier (SiPM) sensors in optical wireless communication (OWC) has shown very promising performance when detecting optical signals with low intensity. However, the recovery period of individual SiPM microcells introduces a nonlinear response and also causes intersymbol interference (ISI) when the transmission data rate is high. The use of a neural network based demodulator has been demonstrated to be more accurate for signal demodulation, especially when the SiPM is nonlinear. Among various neural network algorithms, artificial neural network (ANN) and radial basis function neural network (RBFNN) are two promising candidates. In this paper, we focus on comparing these two neural network structures for signal demodulation. First, to intuitively explain how these two neural networks can be applied for signal demodulation, we consider their structures with a small number of neurons and the constellations of binary phase-shift keying (BPSK). In particular, for ANN, we consider a single neuron so that the signal demodulation step is equivalent to a logistic regression task. In the case of RBFNN, we consider a network with six neurons. By showing the detailed updating process of both networks, we demonstrate that both ANN and RBFNN can achieve almost identical classification boundaries in the Euclidean space. Finally, we consider more complex structures for both ANN and RBFNN, along with larger constellation sizes. The results show that these two structures achieve identical performance for a wide range of irradiances.