Developing Trustworthy Reinforcement Learning Applications for Next-Generation Open Radio Access Networks

Artificial intelligence is envisioned to transform the design and operation of 6G networks. Reinforcement learning (RL), in particular, has emerged as a fundamental approach toward this goal with strong support from the industry and the Open Radio Access Network (O-RAN) Alliance. While research efforts have demonstrated the potential of RL, the lack of trustworthiness of RL algorithms remains a challenge to its adoption in real-world networks. In this paper, we propose a trustworthy RL framework that addresses the core challenges experienced by RL-based radio resource management applications deployed in O-RANs. We then demonstrate a case study on O-RAN slicing incorporating several modules of the proposed framework. The experimental results show improvements in the average RL convergence rate, initial reward value, percentage of converged scenarios, and reward variance. Hence, the RL-based algorithms exhibit fast convergence and enhanced generalizability, safety, and robustness.