Transition Noise Facilitates Interpretability

Ronald Parr, Cynthia Rudin, Harry Chen, Zachery Boner, Michal Moshkovitz, Lesia Semenova

Keywords: Markov Decision Process, Reinforcement Learning, Interpretability

TL;DR: Transition noise can be equivalent to discount, which can short horizons, which can facilitate interpretability

Abstract: Recent research in supervised learning has demonstrated that noise in data generation processes leads to the existence of accurate and simpler/interpretable machine learning models. However, the implications of this effect in the context of reinforcement learning, specifically in Markov Decision Processes (MDPs), have not been thoroughly explored. This paper investigates how noise influences the interpretability of MDPs. For two different types of transition noise, adding noise is provably equivalent to solving a noiseless MDP with a smaller discount factor. Regardless of the value function or policy function representation, problems with shorter planning horizons may be more conducive to interpretable solutions, simply because short-term effects and consequences tend to have more concise representations.

Submission Number: 7