Efficient Deception-Aware Learning of Game-Theoretic Solutions for Adversarial Domains
Abstract
Many strategic interactions today, particularly those in adversarial domains, are rife with uncertainties, information asymmetry and deception. This leads to asymmetric and inaccurate agent knowledge regarding the game environment and their opponents, and thus renders classic game-theoretic analysis built upon the common knowledge assumption inapplicable to such complex strategic games. This project aims to fill this gap by developing data-driven approaches to efficiently learn the optimal strategic decisions. To bridge game-theoretic modeling and data, this project will integrate machine learning, game theory and algorithm design in order to develop provably effective machine learning algorithms that can assist security agencies to intelligently and robustly act in today s complex strategic environments. Specifically, this project has three thrusts which address three fundamental challenges in learning game-theoretic solutions in adversarial environments. The first thrust tackles the challenge of instability of the learner s utility in opponents behaviors. It looks to explicitly model such noise in opponents behaviors --- via either worst-case analysis or boundedly rationality modeling --- and then design learning algorithms that explicitly take such noisy opponent behaviors into account. The second thrust departs from the classics sample-efficient learning framework (which is particularly common for patten-recognition-type tasks), and looks to design algorithms that maximize the learner s accumulated utility in sequential multi-agent settings with unknown strategic opponents. This is motivated particularly by applications in adversarial domains where a single bad decision may be extremely costly, and thus effective learning algorithms must carefully balance the number of sampled plays and their (possibly huge) costs. Finally, the third thrust addresses potential deceptive behaviors from adversaries in order to avoid the learner/defender from being manipulated or misled by the adversary during learning. Notably, such concern of adversaries deceptive behaviors and effective ways to counteract are particularly well-motivated in dversarial domains. Besides developing efficient deception-aware learning algorithms, this thrust will also develop orthogonal methods to combat deceptive adversary behaviors, including actively eliciting information from adversaries under incentive constraints and coping with mis-specified prior knowledge. The expected outcome of this project includes both theoretically grounded learning algorithms as well as open-sourced software which include the developed new algorithms from this project. Besides advancing game-theoretic decision making in strategic settings with uncertainty and asymmetric knowledge, this research will potentially lead to novel machine learning tools to optimize defense strategies in military applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 19, 2023
- Source ID
- W911NF2310030
Entities
People
- Haifeng Xu
Organizations
- Army Contracting Command
- United States Army
- University of Chicago