A New Approach to Tackle Channel Uncertainty with Limited Data Samples

Channel state information (CSI) is a key realtime parameter for resource allocation in wireless networks. However, CSI acquisition or estimation is known to be imperfect and bound to errors and uncertainties. Traditional methods addressing CSI rely on either model-based or model-free approaches, each with its limitations. A model- based approach assumes CSI (or CSI error) follows certain distributions and develops a solution following this assumption. However, the assumed model may not accurately characterize reality, especially as the operating environment changes over time. On the other hand, a model-free approach uses data-driven techniques such as machine learning (ML) for real-time resource allocation. This approach is only effective when a large set of high-quality training data is available. Its convergence is also an issue when there is a limited amount of data set available. Further, this approach does not offer any performance guarantee. In this article, we introduce a new "small-data" approach that only requires a small set of data samples. We show that this approach is able to combine the benefits of both approaches -- offering a performance guarantee solely based on the small set of data samples and without assuming any distributions for the CSI (or CSI error). We will describe how this new approach works and discuss its theoretical underpinning. We further demonstrate the efficacy of this approach by applying it to solve a resource allocation problem involving CSI uncertainty.