Adaptive Mobile Sensing Using Reinforcement Learning Framework

Mobile sensing has created unprecedented opportunities to study human behaviors and serve users in diverse applications. The foundation to mobile sensing is data collection through both passive and active sensing. Successful mobile sensing applications require efficiently managing energy consumption in passive sensing using smartphone embedded sensors, and compliance in active sensing such as mobile Ecological Momentary Assessments (EMAs). To date, there is a lack of a unified framework that can enable adaptive mobile sensing in a personalized and adaptive manner to address both of these challenges. This dissertation leverages the most recognizable general purpose artificial intelligence framework, reinforcement learning (RL), to model both passive and active sensing as sequential control problems, and adapt the sensing tasks to the users' contexts. We design both adaptive passive and active sensing strategies under the RL framework with different problem formulations to improve energy efficiency in passive sensing, and user compliance in active sensing. Performance of the proposed RL strategies are evaluated in simulations using real data collected by continuous mobile sensing in mental health studies. Results from simulations and predictive models show that our approaches, when compared to various baseline methods, consistently achieve: 1) for passive sensing, more energy saving with comparable data utility; and 2) for active sensing, higher overall compliance. We implement and maintain a cross-OS mobile adaptive sensing platform, on which the proposed RL strategies will be evaluated in future studies, and point out future directions to advance mobile sensing technologies.