Learning option MDPs from small data

摘要

Learning from small data is a challenge that presents itself in applications of human-robot interaction (HRI) in the context of pediatric rehabilitation. Discrete models of computation such as an Markov decision process (MDP) can be used to capture the dynamics of HRI, but the parameters of those models are usually unknown and (human) subject dependent. This paper combines an abstraction method for MDPs, with a parameter estimation method originally developed for natural language processing, designed specifically to operate on small data. The combination expedites learning from small data and offers more accurate models that lend themselves to more effective decision-making. Numerical evidence in support of the approach is offered in a comparative study on a small grid-world example.