Leveraging LLMs to eXplain DRL Decisions for Transparent 6G Network Slicing

The emergence of 6G networks heralds a transformative era in network slicing, facilitating tailored service delivery and optimal resource utilization. Despite its promise, network slice optimization heavily relies on Deep Reinforcement Learning (DRL) models, often criticized for their black-box decision-making processes. This paper introduces a novel Composable eXplainable Reinforcement Learning (XRL) framework customized for distributed systems like 6G Network Slicing. The proposed framework leverages Large Language Models (LLMs) and Prompt Engineering techniques to elucidate DRL algorithms’ decision-making mechanisms, with a specific emphasis on user profiles. The latter transforms the inherently opaque nature of DRL into an interpretable textual format accessible not only to eXplainable AI (XAI) experts but also to diverse network slice provider stakeholders, engineers, leaders, and beyond. Experimental results underscore the efficacy of the proposed Composable XRL framework, showcasing substantial improvements in transparency and comprehensibility of DRL decisions within the context of 6G network slicing.