Gates Hillman Center Indoor Location Prediction (GILP)
February 13, 2014
As a consquence of our work in residential, predictive energy consumption we got interested in investigating the use of HCI methods and applied Machine Learning to reduce the consumption of commercial and educational buildings. Through understanding the occupant’s daily spatiotemporal routine and predicting future locations of occupants my goal is to reduce the energy consumption of commercial and educational office buildings.
This work is being conducted in Carngie Mellon University’s (CMU) Gates-Hillman-Center (GHC) in cooperation with CMU’s facility managment services. The GHC offeres us a rich instrumented environment that already tracks office occupancy, temperature, and several key variables associated with Heating, Air Conditioning, and Ventilation. We collected two data sets for the GHC: one temperature/occupancy data set for 298 rooms and Wi-Fi based location tracking data set for 8 office workers. The latter data set was collected with help of a probabilistic algorithm that Yao Dezhong developed for the project. The results of the data collection and the algorithm we used for the energy efficient tracking are published in Future Generation Computer Systems 2013.
A first exploration into the space of indoor location prediction resulted in Indoor-ALPS: An Adaptive Location Prediction System. Indoor-ALPS uses three high level concepts to make predictions based on past location traces: continuous feature selection using SFFS, ensemble prediction, and incremental learning. Using 10 different spatiotemporal features that in my expertise best model a person’s spatiotemporal structure the algorithm solved three different prediction tasks: when will a person transition, where will they transition to, and the combination of both. Since human behavior is highly variable and can change over time Indoor-ALPS is re-evaluating the prediction model after each day by using SFFS to compute a new feature set that best describes the historic spatiotemporal data. Test showed that our algorithm significantly improves over the baseline ZeroR and previous work. The results are published in UbiComp 2014.
We’re currently preparing a journal paper that reports on an extensive comparison of three state-of-the-art indoor location prediction techniques: Adaptive Confidence Estimator, Prediction by Partial Match, and PreHeat. All three algorithms were tested on four data sets, two of them are publicly available (the Augsburg Indoor Location Tracking Benchmark and the Wireless Topology Discovery data set) and the occupancy dataset we collected in the GHC.