ReLVaaS: Verification-as-a-Service to Analyze Trustworthiness of RL-based Solutions in 6G Networks
Xin Tao, Irene Vilà, Swarup Kumar Mohalik, J. Pérez-Romero, O. Sallent · 2025
Reinforcement learning (RL) based solutions are emerging as key enablers for integrating Artificial Intelligence (AI) in future 6G networks due to their ability to tackle complex decision-making and control problems. However, current testing-based validation and evaluation methods for these models may not guarantee their trustworthiness, thereby hindering their adoption in real-world applications, especially for safety or mission-critical scenarios. In contrast, formal verification is presented as a promising solution, but its effective use in RL verification requires key steps of property specification, effective system model construction, and framework support for different analyses. Towards this, we present ReLVaaS, Reinforcement Learning Verification-as-a-Service. ReLVaaS provides templates to facilitate the specification of complex trustworthiness properties, requirement-guided system model construction, and invoking off-the-shelf model checking engines, all in an automated process. ReLVaaS is demonstrated on a DRL-based (Deep Reinforcement Learning) capacity sharing solution for Radio Access Network (RAN) slicing, showcasing the capabilities of probabilistic model checking in providing quantitative analysis of trustworthiness properties of resilience, safety, and reliability.