Towards Trustworthy Reinforcement Learning-based Resource Management in Beyond 5G

It is envisaged that future Beyond 5G (B5G) systems will make extensive use of Artificial Intelligence (AI) capabilities to achieve an efficient automated management and optimization of communication and computing resources and to support advanced data-driven applications that provide the users with highly immersive services. Ensuring the trustworthiness of the AI solutions is key for their successful introduction in B5G, as this will guarantee their robustness towards errors and potential attack threats, the privacy of the used data and trained models and the explainability of AI-based decisions, ensuring that they do not have unsafe consequences. In this context, this paper focuses on the trustworthiness of Reinforcement Learning (RL) solutions, as their inherent trial-and-error behavior during training makes them particularly challenging from the robustness perspective. Then, the paper proposes a framework for managing the lifecycle of an RL-based resource management solution for both training and inference stages to ensure its trustworthy operation. The framework relies on an RL training configuration function to specify the training conditions, a Network Digital Twin (NDT) to perform the training on a safe environment and a continuous operation function to monitor the behaviour of the trained policy during inference. The framework is illustrated with an applicability use case of capacity sharing for network slicing.