Performance comparison of various end-to-end learning technologies with a bandwidth-limited OWC system

Area of Science:

• Optical Wireless Communication (OWC)
• 6G Native Intelligent Systems
• End-to-End (E2E) Learning

Background:

• E2E learning is a promising approach for optimal communication in 6G intelligent systems.
• A lack of comparative analysis hinders the understanding of E2E techniques' applicability and effectiveness.
• Optical wireless communication (OWC) systems face challenges in bandwidth limitations and environmental turbulence.

Purpose of the Study:

• To comprehensively compare three advanced E2E methods: autoencoder-based geometric shaping (AEGS), comprehensive autoencoder (CAE), and wave-wise auto-equalization (WWAE).
• To evaluate their performance in a real bandwidth-limited OWC system.
• To introduce a novel attention-based comprehensive noise joint channel estimator (ACNJCE) as a universal channel model.

Main Methods:

• Implementation and comparison of AEGS, CAE, and WWAE models within a real OWC system.
• Utilizing a novel ACNJCE for adaptable channel modeling.
• Testing under carrier-less amplitude and phase modulation (CAP) and bit-power loading discrete multi-tone modulation (DMT) with varying baud rates (2-3 GBaud).

Main Results:

• CAE demonstrated autonomous bandwidth allocation and a 69% higher dynamic adjustment range than NN-equalized CAP.
• AEGS showed significant received optical power (ROP) gain and superior, stable performance under turbulence.
• WWAE with DMT achieved the highest data rate (near 1 GHz bandwidth) with a BER 25.2% of NN equalization.

Conclusions:

• The study provides the first comprehensive evaluation of major E2E algorithms in a real OWC system.
• CAE and AEGS offer distinct advantages in bandwidth allocation and ROP gain, respectively.
• WWAE with DMT is effective for high data rates in spectrally constrained OWC systems, with AEGS showing resilience to turbulence.