Understanding Reinforcement Learning Based Localisation as a Probabilistic Inference Algorithm

Abstract

Indoor localisation and tracking in a residential home setting are envisaged to play an essential role in smart home environments. As it is hard to obtain a large number of labelled data, semi-supervised learning with Reinforcement Learning is considered in this paper. We extend the Reinforcement Learning approach, and propose a reward function that provides a clear interpretation and defines an objective function of the Reinforcement Learning. Our interpretable reward allows us to extend the model to incorporate multiple sources of information. We also provide a connection between our approach and a conventional inference algorithm for Conditional Random Field, Hidden Markov Model and Maximum Entropy Markov Model. The developed framework shows that our approach benefits over the conventional algorithms a real-time prediction scenarios. The proposed Reinforcement Learning method is compared against other supervised learning approaches. The results suggest that our method can learn in the semi-supervised learning setting and performs well in a small labelled data regime.

This work was supported by SPHERE Next Steps Project funded by the U.K. Engineering, and Physical Sciences Research Council (EPSRC) under Grant EP/R005273/1 and the UKRI Turing AI Fellowship EP/V024817/1.