Learning-Aided Unary Error Correction Codes for Non-Stationary and Unknown Sources

Abstract

Unary error correction (UEC) codes have recently been proposed for the joint source and channel coding of symbol values that are selected from a set having an infinite cardinality. However, the original UEC scheme requires the knowledge of the source probability distribution, in order to achieve near-capacity operation. This limits the applicability of the UEC scheme, since the source probability distribution is typically non-stationary and is unknown in practice. In this paper, we propose a dynamic version of the UEC scheme, which can learn the unknown source statistics and gradually improve its decoding performance during a transient phase, then dynamically adapt to the non-stationary statistics and maintain reliable near-capacity operation during a steady-state phase, at the cost of only a moderate memory requirement at the decoder. Based on the same learning technique, we also propose two separate source and channel coding benchmarkers, namely, a learning-aided Elias gamma-convolutional code (CC) scheme and a learning-aided arithmetic-CC scheme. The simulation results reveal that our proposed learning-aided UEC scheme outperforms the benchmarkers by up to 0.85 dB, without requiring any additional decoding complexity or any additional transmission-energy,-bandwidth, or-duration.