Neural Network-Assisted Robust Symbol Detection Under Intersymbol Interference

In recent years, the machine learning assisted communication system design has drawn a lot of attentions. As a remarkable progress, a recent work proposed to incorporate a neural network (NN) into the traditional algorithms for symbol detection under intersymbol interference (ISI), e.g. the Viterbi algorithm and the BCJR algorithm, to achieve robustness against channel estimation errors. This paper presents an improved design over the state-of-the-art by using a neural network to approximate the likelihood of the received sample given different state transitions of the trellis diagram. The simulation results show that the proposed method performs similarly to the conventional methods in the channel model-matched scenarios, but is significantly more robust against channel estimation errors. Our design is superior to the state-of-art NN -assisted methods in two aspects: it requires significantly smaller training overhead and is robust against non-Gaussian noise.