Approximate dynamic programming

Project Summary

Intelligent mobile sensor agent can adapt to heterogeneous environmental conditions, to achieve the optimal performance, such as demining, maneuvering target tracking. 

Themes and Categories
Year
Contact
Wenjie Lu
wenjie.lu@duke.edu

The mobile sensor agent is a robot with onboard sensors, and it is deployed to navigate obstacle-populated workspaces subject to sensing objectives. The expected performance of available future measurements is estimated using information theoretic metrics, and is optimized while minimizing the cost of operating the sensors, including distance. Approximate dynamic programming and non-parametric Bayesian models are studied in the heterogeneous system.

Related People

Related Projects

This project is also part of Duke’s first Climate+ cohort

A team of students led by researchers at Duke and abroad will develop and evaluate machine learning solutions to model behavioral patterns of electric use, emphasizing data privacy. Data collected in different parts of the world will be analyzed to understand the electric patterns that characterize various appliances and how that information can model users' consumption profiles and prevent fraud. 

 

Project Lead: J. Matias Di Martino

Project Manager: Raza Lamb

Data+ students led by Prof. Henri Gavin will develop AI models for on-site earthquake early warning, in which sensors at a site provide warnings at that site. The Data+ project will integrate into ongoing work on geophone sensors, IOT microcontrollers, and networking. The Data+ team will focus on machine learning aspects of the project by making use of extensive seismic databases. The machine learning model, an anisotropic Gaussian Process, will relate waveform characteristics from the initial moments of shaking to predict the maximum expected shaking at the monitored site.

 

Project Lead: Henri Gavin

Project Manager: James Matuk, Ph.D.

 

 

 

This project is also part of Duke’s first Climate+ cohort

A student team working with the Energy Data Analytics Lab will work to democratize access to data relevant to climate change mitigation and adaptation planning as well as the underlying models to acquire those data. This project will work towards building the first “foundation model” specifically for remote sensing imagery for the purpose of extracting climate change relevant content at scale to enable near real-time tracking of climate causes and impacts.  A foundation model is a model (usually a deep neural network) that has been trained on a large and diverse set of data, after which it can be adapted to a variety of different (but related) inference tasks with a small fraction additional training data and computation. Leveraging recent developments in self-supervised learning, we will develop the dataset for creating this foundation model and begin training it on real-world data. A model developed using such a dataset will enhance climate change mitigation/adaptation monitoring and planning through developing robust features that can be used to monitor a broad range of climate change contributing factors (e.g. energy infrastructure and use, agricultural activity) and impacts (e.g. economic impacts and human migration) for informing climate mitigation and adaptation strategies.

 

Project Lead: Kyle Bradbury

Project Manager: Zachary Calhoun