Hybrid ensemble learning for triggering of GPS in long-term tracking applications

Abstract

Long-term tracking is an expanding field with applications in logistics, ecology and wearable computing. The main challenge for longevity of tracking applications is the high energy consumption of GPS, which has been addressed by using low power sensors to trigger GPS activation upon detecting events of interest. While triggering can reduce power consumption, static thresholds can under-perform in the long-term as context changes. This paper presents a comparison between a dynamic adaptive threshold algorithm and off-line machine learning techniques. We test the algorithms on empirical data from flying foxes to show that off-line machine learning techniques improve the hit rate when compared to the dynamic adaptive threshold algorithm. We then combine the models into an on/off-line hybrid ensemble learning model to improve both hit rate and false alarm rate when compared to the dynamic adaptive threshold algorithm. The hybrid model also has lower false alarm rate and precision when compared to the stand alone machine learning algorithms. We also test the off-line machine learning techniques on unknown data to show that the hit and false alarm rates vary from node to node. This indicates that more consistent performance might be found through the development of on-line machine learning algorithms.