Representation-Based Continual Learning for Channel Estimation in Dynamic Wireless Environments

Most AI-based channel estimation methods with static environment assumptions suffer from performance degradation due to the distribution shift caused by the varying channel environment. As one of the solutions, transfer learning also faces the problem of catastrophic forgetting, where the model tents to fail in previous estimation tasks after learning from new ones. In this paper, we propose a continuous learning-based channel estimation (CLCE) scheme that integrates a series of subnetworks to preserve historical knowledge and achieve an ongoing process of self-improvement. To determine whether the wireless environment is previously unobserved, we first propose a distance-based unsupervised out-of-distribution (OOD) detection algorithm to perceive the distribution shift of the channel environment. The OOD detection algorithm is developed based on the representation of channel data in the latent space of a variational autoencoder (VAE), which is designed to infer the latent variable that implies the characteristics of the wireless environment. Then, a new channel estimation subnetwork is initiated with meta-learning to adapt to the dynamic channel environments with a small set of OOD channel data. Simulation results reveal that our proposed scheme can accurately detect unobserved channel environments without introducing additional detection network, and efficiently adapt to them with few online samples. Furthermore, the mean square error (MSE) result of the channel estimations across multiple environments demonstrates that CLCE effectively mitigates catastrophic forgetting and outperforms the competitors.