Explainable Reinforcement Learning for Network Management via Surrogate Model

In recent years, the adoption of Machine Learning, particularly Reinforcement Learning (RL), for the control and management of communication networks has emerged. However, a critical challenge hindering its practical implementation is the lack of explainability inherent in these models, which prevents network administrators from adopting these techniques despite their great potential to improve network performance. This paper aims to enhance the trustworthiness of RL-based network management and control, by making the RL model explainable and providing administrators with transparent insights into the RL decision-making processes. With this aim, we propose a methodology that leverages surrogate models, specifically, Decision Trees (DTs), to create simplified yet interpretable representations of the original RL model, able to explain it. Experiments were conducted to evaluate the efficacy of our method, demonstrating that the surrogate model achieves about $94 \%$ accuracy in imitating the original RL model. Additionally, the surrogate model significantly improves the explainability of the entire system by automatically generating graphical representations, in the form of DTs for interpreting the RL decisions.