Evaluation of Translation Layers for Deep-Learning Channel Modelling

Communication systems aim to send messages, from a transmitter to a receiver over a physical channel, with minimal error. However, the signals from the transmitter travel over a communication channel containing unwanted effects that corrupt the signal, making it difficult to reconstruct the original message at the receiver. It is crucial to have a representative model of the real channel during the design phase of communication systems to ensure optimal performance. As channel modelling becomes more complex with evolving scenarios, conventional methods can become increasingly time-consuming, complex, and inflexible. In recent years, deep learning has emerged as a promising alternative for channel modelling, but capturing channel artifacts efficiently, especially under changing conditions, remains a challenge. Translation layers, (or domain translation techniques) in machine learning, offer a solution by adapting models to new environments more rapidly. In this paper, we review existing literature and highlight the limitations of previous studies in channel modelling. We propose an alternative method that integrates a conventional channel model with a deep-learning-based translation layer. Our findings show that this approach preserves the efficiency of the deep-learning-based channel model without significantly affecting convergence time.