Multilayer Extreme Learning Machine as Equalizer in OFDM-based Radio-over-fiber Systems

Abstract

Mobile/wireless networks aim to support diverse services with numerous and sophisticated requirements, such as energy efficiency, spectral efficiency, negligible latency, robustness against time and frequency selective channels, low hardware complexity, among others. From the central station to the base stations, radio-over-fiber orthogonal frequency division multiplexing (RoF-OFDM) schemes with direct-detection are then implemented. Unfortunately, laser phase noise, chromatic fiber dispersion, and carrier frequency offset impair the orthogonality of the subcarriers; hence, deteriorating the performance of the RoF-OFDM system. In order to take all the processing tasks to the cognitive level (the last goal in the telecommunication industry), various extreme learning machines (ELMs), composed by only a single hidden layer, have been recently adopted as equalizers. The reason behind this trend comes from the lower computational complexity, higher detection accuracy, and minimum human intervention of the ELM algorithms. In this article, we introduce a multilayer ELM-based receiver for RoF schemes transmitting phase-correlated OFDM signals affected by phase and frequency errors. Results report that by appropriately setting the hyper-parameters of the multilayer ELMs, the ELM with 3 hidden layers outperforms most of the ELMs reported in the literature (the ELM with 2 hidden layers, original ELM, regularized ELM, and 2 fully-independent ELMs defined in the real domain), as well as the benchmark pilot-assisted equalizer in terms of bit error rate. Nevertheless, this benefit comes with excessive computational cost. Finally, we show that the fully-complex ELM is still the best equalizer taking into account several key metrics.