Unsupervised and Generalizable Wireless Positioning via Variational Reinforcement Learning

Achieving high-precision positioning holds paramount significance in the realms of intelligent manufacturing, the Internet of Things, and smart transportation. Nevertheless, attaining accurate labels and ensuring network generalization on mobile devices poses a formidable challenge. In this paper, we put forth an unsupervised and versatile wireless positioning methodology rooted in deep variational reinforcement learning (DVRL). The primary contributions of this work encompass the following: 1) the introduction of an approach to model wireless localization based on user-equipment (UE) as a partially observable Markov decision process (POMDP) and the treatment of this model within the DVRL framework; 2) the incorporation of a neural processes regressor (NPR) to derive ranging results from environmental cues, thus negating the need for model retraining; and 3) the alleviation of the challenge posed by the need for precise labels by deriving rewards from NPR's localization uncertainty, making use of unlabeled data. Our simulations affirm that this approach yields high-precision positioning without mandating precise labels or network retraining, surpassing conventional techniques in both positioning accuracy and generalization capabilities.