Wireless Channel Prediction using Artificial Intelligence with constrained Data Sets

The work in this paper deals with the study of artificial intelligence (AI) tools for purposes of vehicular wireless channel prediction. The objective is to test the ability of different types of AI algorithms under different types of implementation constraints that will be typical of vehicular applications. We focus particularly in scenarios where the channel changes relatively fast and therefore the relevant measurements or long-term statistical models are scarce. This means that the training of our models can be potentially inaccurate or incomplete and we need to investigate which AI algorithm behaves better in this challenging condition. In future work we aim to investigate also computation complexity constraints, real-time deadlines, and outdated/distorted or noisy data set samples. We also aim to correlate the main properties of the well-known Jakes’ channel model with the effectiveness of the type of prediction and the parameters of the different algorithms being tested. The objective of channel prediction in vehicular networks is to reduce allocation and transmission errors, thereby reducing latency and improving message transmission reliability, which is crucial for future applications such as autonomous vehicles. Results show that even in situations with incomplete data sets, AI can provide good approximate predictions on the channel outcome. LSTM in multi-step and CNN in single-step between implemented algorithms have better performance and lower error in prediction.