New Method Identifies Probabilistic Causes in MDPs Efficiently

Researchers have developed a novel, sample-efficient method for identifying probabilistic causes within unknown Markov Decision Processes (MDPs), offering probabilistic guarantees. This addresses a limitation in traditional probabilistic model checking, which often explains *what* happens but not *why* undesired outcomes occur. The core innovation involves a restart-based modification of the MDP, which simplifies the complex task of checking for "probability-raising (PR) causality." This modification allows the identification of PR causes to be reduced to two conditional reachability queries, crucially without requiring prior knowledge of the original MDP's reachability probabilities. This is particularly beneficial when transition probabilities are unknown, making it suitable for learning scenarios. The paper proves the correctness of this approach, establishes sample-complexity bounds, and presents an anytime learning-and-checking algorithm. Experimental results on benchmark datasets demonstrate that the method reliably and quickly classifies states as causal, non-causal, or undecided, providing valuable insights into system behavior.