Semi-Markovian User State Estimation and Policy Optimization for Energy Efficient Mobile Sensing

User context monitoring using sensors on mobile devices benefits end-users unobtrusively and in real-time by providing information support to various kinds of mobile applications and services. A pervasive question, however, is how the sensors on a mobile device could be scheduled more efficiently such that they can detect more user information with as little energy usage as possible. In this paper, we model the user state sensing problem as the intermittent sampling of a semi-Markov process. Assuming sensors make perfect detections, we propose (a) a semi-Markovian state estimation mechanism that selects the most likely user state while observations are missing, and (b) a tractable approximation to a semi-Markov optimal sensing policy. The approximate semi-Markov optimal sensing policy us determines the duration the sensor should stay idle based on its current state detection before making the next observation, such that the expected state estimation error is minimized while maintaining a given energy budget. Both the semi-Markov estimation mechanism and the novel sensing policy us we develop are evaluated on simulated two-state processes and real user state traces, and their performances are shown to outperform Markov estimation and the Markov-optimal policy (studied in [25]), where the gain depends on the particular state distributions. Finally, we implement an energy efficient human activity recognition system on Nokia N95 smart phones, where the accelerometer is operated according to us . We demonstrate by a set of real experiments that the battery lifetime could be increased to at least three times as compared to continuous sensing, while producing only less than 3% estimation error.