Computational complexity comparison of feedforward/radial basis function/recurrent neural network-based equalizer for...

Area of Science:

• Optical communications
• Signal processing
• Artificial intelligence

Background:

• High-speed optical links like 50-Gb/s pulse amplitude modulation (PAM)-4 direct-detection (DD) systems require advanced equalization techniques to mitigate signal impairments.
• Neural network (NN)-based nonlinear equalizers are emerging as a promising solution for improving bit-error-rate (BER) performance.

Purpose of the Study:

• To analyze and compare the computational complexity and BER performance of four types of NN-based nonlinear equalizers.
• To provide guidelines for selecting appropriate NN-based equalizers for optical communication systems based on specific requirements.

Main Methods:

• Analysis of computational complexity and BER performance for feedforward neural networks (F-NN), radial basis function neural networks (RBF-NN), auto-regressive recurrent neural networks (AR-RNN), and layer-recurrent neural networks (L-RNN).
• Numerical simulations were conducted for a 50-Gb/s PAM-4 DD optical link.

Main Results:

• AR-RNN-based equalizers demonstrated the lowest computational complexity for a fixed BER threshold.
• Amongst equalizers with identical input and hidden neuron counts, F-NNs had the lowest complexity, while AR-RNNs achieved the best BER performance.
• RBF-NNs require more hidden neurons with increasing input numbers, limiting their suitability for long-distance transmissions.
• NN-based equalizers require only tens of multiplications per symbol for good BER performance, enabling real-time implementation.

Conclusions:

• AR-RNNs present an optimal balance of low computational complexity and superior BER performance for 50-Gb/s PAM-4 DD optical links.
• The study offers valuable insights for engineers to select NN-based equalizers tailored to specific BER and computational constraints in optical systems.