Stimulating and Manipulating AI Agents via Recognizant, Tiered Modeling in Adversarial Environments

Interdisciplinary Areas: Data and Engineering Applications, Autonomous and Connected Systems, Smart City, Infrastructure, Transportation

Project Description

Decision making under uncertainty in situations that involve multiple strategic and self-interested participants is a classical problem. The particular application area we will concentrate on is urban transportation systems, both road transport with vehicle fleets (with possibly autonomous vehicles) owned and operated by different companies and air mobility systems with drone delivery from fleets owned by different decision makers. Existing frameworks in this area tend to make assumptions that are too simplified in an age where many of these individual participants will be AI agents / algorithms. Assumptions fundamental to the existing frameworks such as shared knowledge among participants (e.g., on possible strategies and conventions), clearly identified allies and adversaries, and limited strategy space for the participants to optimize over will quickly break down. A fundamentally new approach will be needed to learn behavior/interaction models of these AI agents/algorithms, to actively manipulate them, and to optimize the achieved outcomes. This project provides a framework and coherent plan to this end. We will demonstrate the theoretical advancements on simulation testbeds for the two physical applications mentioned above.

Start Date

February 2025

Post Doc Qualifications

The ideal postdoc candidate would have experience in theoretical machine learning/reinforcement learning and game theory algorithms. The candidate will ideally have PhD in CS, EE, Stats, or related areas and have expertise in algorithms and analysis of guarantees for the algorithms.

Co-Advisors

Vaneet Aggarwal, vaneet@purdue.edu, IE/ECE, https://engineering.purdue.edu/CLANLabs
Vijay Gupta, gupta869@purdue.edu, ECE, https://engineering.purdue.edu/Vijay-Gupta

Bibliography

https://openreview.net/pdf?id=HCl41wIi9gc
https://arxiv.org/pdf/2403.11925
https://arxiv.org/pdf/2305.12633
https://ieeexplore.ieee.org/abstract/document/10549785/
https://arxiv.org/pdf/2402.08747