Research Overview

My research is to bridge Cyber Physical Systems (CPS), Data Science and Applied Machine Learning in extreme-scale urban infrastructure from a sociotechnical perspective. In the last 10 years, my group has been focused on the life cycle of CPS from mobile sensing to uncertainty-oriented prediction to socially-informed decision-making, by exploring fundamental and use-inspired CPS science. Our application domains include transportation (e.g., taxis, buses, subways, bikes, cars, electric vehicles), telecommunication (e.g., cellphones and Wi-Fi), payment (e.g., smartcards), social networks (e.g., check-in), and sharing economy (e.g., on-demand delivery) across 8 cities on 3 continents with 500 thousand vehicles, 10 million phones, 16 million smartcards, and 100 million residents involved. Our intellectual core is \textit{trustworthiness} on mutually-beneficial Human-CPS partnerships via their synergistic interaction focusing on real-time human mobility.

Research Framework

As in the left figure, we focus on a 3-layer closed-loop framework for mutually-beneficial human CPS interactions. Its key novelty is the cross-domain philosophy where these cross-domain infrastructures and data complementarily model physical phenomena and human behavior for a synergistic Human-CPS partnership.

Layer 1: Mobility-Centric Sensing

We work with both industry and government (e.g., Alibaba and City of Newark NJ) to utilize existing and deploy new sensing systems at scale, e.g., Nationwide indoor localization deployment aBeacon. These sensors include GPS, camera, Bluetooth, Wi-Fi, RFID, and humans as sensors for crowdsourcing. These systems produce real-time cross-domain data for prediction models to forecast physical and behavioral measures.

Layer 2: Uncertainty-Oriented Prediction

Based on cross-domain data from the lower-layer sensing, we have been exploring various measures to quantify and predict phenomena of interest (e.g., mobility) in terms of locations, flows, time, speeds, energy, modes, routes, and privacy by different statistical and learning models. Most of these measures are highly uncertain given different contexts, which require new techniques based on Transfer Learning, Federated Learning, Generative Adversarial Networks, Recurrent, Convolutional, and Graphical Neural Networks. These prediction models will provide cross-domain knowledge for the top layer decision-making models.

Layer 3: Socially-Informed Decision-Making

Based on predicted measures, we have been designing a few services for the decision-making layer including ride-sharing, car-sharing, on-demand delivery, travel and parking recommendations, electric vehicle charging, vehicle dispatching, transit transfer, advertising, based on various Control (e.g., Receding Horizon Control), Learning techniques (e.g., Deep Reinforcement Learning) and Game Theory (e.g., Multi-agent Stochastic Game) to social and behavioral factors. These services provide positive cross-domain feedback to improve the physical world and humans.

Based on various metrics for trustworthiness as in the left figure, a closed-loop optimization is created where the improved physical world and humans generate new and positive signals for sensing systems to produce new data, which are obtained by prediction models to produce new and better knowledge for decision-making for new feedback to further improve the physical world and humans, i.e., forming a synergistic Human-CPS partnership.