Jia Liu

Collaborative Research: CPS: Medium: Real-Time Crowd-Sourced Geospatial Digital Twin for Cyber-Physical Systems

NSF 2331104

Lead PI: Jia (Kevin) Liu

Co-PIs: Bin Li (Penn State), Randall Berry (Northwestern), and Rongjun Qin (Ohio State)

Intellectual Merits

This research project focuses on enhancing the way vital information is delivered to smart mobile devices—such as smartphones and tablets. With the advancement of technology, there is a growing necessity for these devices to receive various types of information (like images, videos, and texts) instantly and effectively. One promising approach to achieving this is through the use of Geospatial Digital Twins (GDT), which are digital models of physical environments. GDTs are becoming increasingly important as they allow for real-time updates and interactions, making them invaluable for various applications such as monitoring, maintenance, and emergency responses. Traditionally, data for GDTs has been collected through automated systems like satellites and drones. However, these methods have limitations, especially when it comes to updating data quickly and covering hard-to-reach areas. To overcome these challenges, this project is developing a novel approach that involves the community through “human-in-the-loop” strategies. This means using crowd-sourced data, where people on the ground provide real-time updates to digital models. This method not only enhances the accuracy and timeliness of the information but also discovers new information. The project has the potential to revolutionize how we interact with and understand our physical world, making it a cornerstone for further scientific and educational advancements. It also plays a vital role in education, integrating research findings into university curricula and offering unique learning opportunities for students, including those from underrepresented backgrounds.

Major Activities

Ensuring Crowd-SourcedData Freshness for GDT
In our first thrust, we adopt the notion of Age-of-Information (AoI) as a measure of information freshness. We aim to develop efficient incentive mechanisms to guarantee the freshness of collected imagery and meta data while minimizing the information mismatch between our GDT system and its physical counterpart. We will first study the optimal trade-off between the system performance and reward rate. Then, we will address the random agent behaviors. Finally, we will ensure the system operates properly under limited budgets.
Integrating Crowd-Sourced Data for Real-Time GDT Update
In our second thrust, we focus on developing robust methods that leverage crowd-sourced data collected by users to update GDT where needed. Both agent- and cloud-level algorithms will be developed to guide the users to collect quality data, and seamlessly integrate these data into GDT for consistent updating. The results will establish a common testbed where data freshness, truthness and equity are considered when updating.
Guaranteeing Truthful Reporting in Crowd-Sourced Data Collection
In our third thrust, we focus on incentive mechanisms for another key aspect of crowd-sourced real-time data collection -- guaranteeing truthful and accurate reporting. We will develop game-theoretic models for the interaction between the crowd-sourcing platform and use this to develop mechanisms to incentivize users to submit truthful and accurate reports. We will further use this framework to address a number of important real-world challenges in truthful reporting, such as imperfect information and user heterogeneity.
Mitigating Self-Reinforcing Bias in Crowd-Sourced GDT Update
Our last thrust will mitigate the self-reinforcing user preference effect in crowd-sourcing. Specifically, we propose a new three-phase incentivized bandit-learning policy that mitigates the users' initial and self-reinforcing bias and the self-serving strategic behaviors at each point of interest (PoI), which allows one to achieve high learning accuracy and efficiency with low incentive costs.

Products

Zhiyao Zhang, Myeung Suk Oh, FNU Hairi, Ziyue Luo, Alvaro Velasquez, and Jia Liu, "Finite-Time Global Optimality Convergence in Deep Neural Actor-Critic Methods for Decentralized Multi-Agent Reinforcement Learning," in Proc. ICML, Vancouver, Canada, Jul. 2025 (acceptance rate: 26.9%).

Jia (Kevin) Liu,