A Novel Hybrid Interpretability Method for Sequential Decision Making

Abstract

We consider the problem of interpretability for sequential decision making, which is frequently addressed by the Markov Decision Processes (MDPs) approach. We distinguish interpretability toward two types: (i) for a machine and (ii) for humans.The key difference between the two is that interpretability for a machine helps simplify the model and for humans helps understand the recommendations. We propose a hybrid approach combining these two types of interpretabilities to achieve better user satisfaction; for this we utilized i) the Logical NeuralNetwork (LNN) and ii) the classical Decision Tree (DT) techniques.