Abstract

Underwater acoustic (UWA) sensor and control links carry mostly short messages (below ∼10 kB) over time-varying, low-SNR multipath channels, which is a regime where forward error correction (FEC) operates on short, finite blocks where a single static configuration is inefficient. Adaptive schemes for these links typically adjust the modulation order and code rate; the payload segment (block) size—which, together with the code rate, sets the coded block length that governs the finite-blocklength penalty for short messages—is seldom adapted per transmission jointly with the choice of FEC scheme on a like-for-like footing. We propose a per-transmission controller that jointly selects the FEC scheme, code rate, and segment size from a prediction of the near-term channel state, which is paired with a like-for-like short-block benchmark of LDPC, list-decoded polar, BCH, Reed–Solomon, convolutional, and turbo codes. No single code dominates: under a unified ARQ goodput metric, the reliability–throughput frontier has a crossover that shifts with the channel, so the optimal FEC choice is channel-dependent. Across our experiments, the segment-size degree of freedom is the dominant throughput lever, capturing essentially all of the adaptation gain at low-to-mid SNR and over fading; switching the FEC family adds a further, bounded gain only where the frontier crosses between families (up to 10% at high-SNR AWGN, polar to RS). The joint controller essentially matches a fair single-family adaptive baseline off the crossover (to within a negligible prediction-overhead margin) and exceeds it at that crossover, beats a fixed, no-CSI code by up to 16%, and captures 91–99% of an oracle; a lightweight persistence predictor matches a learned LSTM for the first-order channel-state model studied. A statistics-driven replay using measured-channel parameters, and a recorded-channel replay over the public Watermark benchmark, preserve the same family ordering.

IPC Classification

Keywords