Learning Without Forgetting: Predicting the Reliability of V2X Wireless Communication

Effective communication between vehicles and road users is essential for reducing accidents and congestion. Reli-able wireless communication is crucial for decision-making in advanced driver assistance systems and autonomous vehicles. In this work, we propose a convolutional neural network to predict the frame error rate in vehicle-to-infrastructure scenarios. Using a geometry-based stochastic channel model and hardware-in-the-loop emulation, we generate a dataset on which our model achieves 90 % validation accuracy. To adapt the model to new data, such as vehicle-to-vehicle scenarios, and to reduce computational costs for retraining the entire model from scratch, we explore methods like fine-tuning, transfer learning, and learning without forgetting (LwF). While these methods improve performance on new data, they reduce accuracy on the original data. To address this, we modify LwF by including some original data, achieving a balanced accuracy of 81.96%.