This paper investigates the problem of inferring knowledge from data so that the inferred knowledge is interpretable and informative to humans who have prior knowledge. Given a dataset as a collection of system trajectories, we infer parametric linear temporal logic (pLTL) formulas that are informative and satisfied by the trajectories in the dataset with high probability. The informativeness of the inferred formula is measured by the information gain with respect to given prior knowledge represented by a prior probability distribution. We first present two algorithms to compute the information gain with a focus on two types of prior probability distributions: stationary probability distributions and probability distributions expressed by discrete time Markov chains. Then we provide a method to solve the inference problem for a subset of pLTL formulas with polynomial time complexity with respect to the number of Boolean connectives in the formula. We provide implementations of the proposed approach on explaining anomalous patterns, patterns changes and explaining the policies of Markov decision processes.
تحميل البحث