CPS Small Human in the Loop Learning of Complex Events in Uncontrolled Environments

Project: Research project

Project Details

Description

CPS Small Human in the Loop Learning of Complex Events in Uncontrolled Environments CPS: Small: Human-in-the-Loop Learning of Complex Events in Uncontrolled Environments This project aims to build a human-in-the-loop cyber-physical system (CPS) that can be used to provide adaptive automated interventions. The pilot application of the project, which is the focus on the validation approach, is personalized behavioral medicine. In such a system, the state of human behavior is assessed, continuously, using sensor data collected with wearable sensors and mobile devices. The system closes the loop by providing behavioral treatments based on the current state of the user. To build such a cyber-physical system, the project will develop a machine learning approach to model human behavior. However, designing a supervised machine learning model requires acquiring ground truth labels about a persons behavior. Gathering ground truth data in uncontrolled environments is not straightforward because users can introduce various types of noise into the machine learning systems by incorrectly expressing or articulating their behavior. To tackle this challenge, we propose a human-in-the-loop approach for data gathering and machine learning model training. The core computational approaches that this project focuses on include mindful active learning and complex behavior inference. The major goals of the project include (1) mindful active learning: we design active learning strategies that take constraints of the user such as query budget, informativeness of sensor data, and human memory into consideration when deciding what sensor data to query. To this end, the project will design both real-time active learning and offline active learning algorithms; (2) complex behavior inference: the project develops a vocabulary of complex events by analyzing user-expressed structured and unstructured input and uses this knowledge along with the sensor data to construct graph-based learning algorithms for behavior inference; (3) validation: the algorithms and techniques are validated through publicly available data as well as a user study that involves data collection with mobile devices in uncontrolled settings.
StatusActive
Effective start/end date1/1/2212/31/22

Funding

  • National Science Foundation (NSF): $303,891.00

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.