Attribution-based Explanations for Markov Decision Processes

A new research paper submitted to arXiv introduces novel attribution techniques designed specifically for Markov Decision Processes (MDPs), addressing a significant gap in explainable artificial intelligence. While traditional attribution methods focus on static input features at single time points, they often fail in sequential decision-making contexts. This study formally characterizes attributions within MDPs, assigning importance scores to both individual states and execution paths. The authors demonstrate how these scores can be efficiently computed by leveraging strategy synthesis techniques, effectively handling the inherent non-determinism of MDPs. The proposed approach was evaluated across five case studies, proving its utility in providing interpretable insights into the logic of sequential decision-making agents. This work contributes to the field of artificial intelligence by enhancing the transparency and interpretability of complex AI models used in dynamic environments, offering a robust framework for understanding agent behavior over time.