Predicting Link Failures With Online Meta-Learning Under Time-Varying Blockage Dynamics

The increasing demand for higher communication frequencies, due to bandwidth limitations, has highlighted the issue of loss from high-frequency electromagnetic wave blockages. This loss significantly impacts communication performance, making accurate prediction of link failures from line-of-sight (LoS) path blockages essential for minimizing power wastage. Traditional offline link failure prediction methods require extensive memory for historical data storage at base stations and substantial computational resources to develop predictors adaptive to rapidly changing environments. To address these challenges, we employed the follow the meta leader (FTML) algorithm from online meta-learning, enabling quick adaptation to time-varying blockage dynamics with minimal data. We also introduce round reservoir sampling, an efficient storage management technique, to optimize memory usage under constraints. Experimental results show that our LSTM-based predictor, using this online meta-learning approach, achieves faster adaptation and higher accuracy with less data on new blockage dynamics under memory-constrained conditions, compared to other baseline methods.