Neural network decoder of polar codes with tanh-based modified LLR over FSO turbulence channel

Area of Science:

• Optical Communications
• Coding Theory
• Machine Learning

Background:

• Free-space optical (FSO) communication systems are susceptible to atmospheric turbulence, degrading signal quality.
• Polar codes offer efficient error correction but require complex decoders.
• Deep learning approaches are emerging for decoding complex codes.

Purpose of the Study:

• To investigate the first deep learning-based decoder for polar codes specifically over FSO turbulence channels.
• To design and train a neural network (NN) decoder capable of learning polar code encoding rules and channel characteristics.
• To evaluate the performance and stability of the NN decoder under various turbulence conditions.

Main Methods:

• Utilized feedforward neural networks (NN) to construct the polar code decoder.
• Designed custom NN layers for training within the turbulence channel environment.
• Proposed a tanh-based modified log-likelihood ratio (LLR) as the input to the NN decoder, comparing it with the standard LLR.
• Conducted simulations to assess bit error rate (BER) performance.

Main Results:

• The NN decoder with tanh-based modified LLR input demonstrated faster convergence and superior BER performance compared to standard LLR.
• The proposed NN decoder achieved BER performance comparable to conventional Successive Cancellation List (SCL) decoders in FSO turbulence.
• The NN decoder successfully learned the encoding rules of polar codes and the statistical properties of the turbulence channel.
• The trained NN decoder exhibited stable performance even when tested under different turbulence conditions than those used for training.

Conclusions:

• Deep learning-based decoders, particularly with the proposed tanh-based modified LLR, are effective for polar codes in FSO turbulence channels.
• The NN decoder shows promise for practical FSO communication systems due to its performance and adaptability to channel variations.
• This research validates the capability of neural networks to learn complex coding schemes and channel impairments simultaneously.