Predicting user movements in heterogeneous indoor environments by reservoir computing

Anticipating user localization by making accurate predictions on its indoor movement patterns is a fundamental challenge for achieving higher degrees of personalization and reactivity in smart-home environments. We propose an approach to real-time movement forecasting founding on the efficient Reservoir Computing paradigm, predicting user movements based on streams of Received Signal Strengths collected by wireless motes distributed in the home environment. The ability of the system to generalize its predictive performance to unseen ambient configurations is experimentally assessed in challenging conditions, comprising external test scenarios collected in home environments that are not included in the training set. Experimental results suggest that the system can effectively generalize acquired knowledge to novel smart-home setups, thereby delivering an higher level of personalization while decreasing costs for installation and setup.