My research focuses on developing theoretical foundations and computational methods that enable autonomous systems to operate in uncertain, complex, and potentially adversarial domains with verifiable safety and performance guarantees.
The following are some of the projects on which I have been working during the past few years.
Planning in Adversarial Environments
Consider a mobile robot that aims to complete a mission in a stochastic environment while being observed by adversaries. How can the robot protect its mission-critical information from adversaries with unknown locations and capabilities? How would the robot's imperfect sensor measurements affect its ability to protect the mission-critical information? Would the robot improve its performance if it had some side information regarding the adversaries' perception or prediction capabilities? In this project, I propose information-theoretic measures to rigorously express the robot's objective, establish several theoretical results that characterize the scenarios in which the robot can achieve its objective, and develop efficient algorithms to synthesize strategies that leak minimum information to adversaries regarding the robot's mission and trajectories.
Related Publications:
Entropy Maximization for Markov Decision Processes Under Temporal Logic Constraints, Y. Savas, M. Ornik, M. Cubuktepe, M. O. Karabag, U. Topcu, IEEE Transactions on Automatic Control, 2019 [Link]
Unpredictable Planning Under Partial Observability, M. Hibbard, Y. Savas, B. Wu, T. Tanaka, U. Topcu, IEEE Conference on Decision and Control, 2019 [Link]
Minimizing the Information Leakage Regarding High-Level Task Specifications, M. Hibbard, Y. Savas, Z. Xu, U. Topcu, IFAC World Congress, 2020 [Link]
Sequential Incentive Design
Consider a principal that aims to modify an agent's behavior through a sequence of incentive offers. For example, an online retailer may aim to convince a customer to purchase more products over time by offering a sequence of discounts, or a ridesharing company may aim to convince a driver to be present at a certain location during rush hour by offering additional cash income for a series of rides. Is it possible to modify the agent's behavior if the agent has an unknown intrinsic motivation? If possible, what is the minimum cost of behavior manipulation? What are the memory requirements for the principal to minimize its cost? Does there exist an efficient algorithm to compute an incentive sequence that minimizes the cost to the principal? In this project, I establish several theoretical results that characterize the computational complexity of the considered behavior manipulation problem and illustrate the value of memory for the principal. I also propose several tractable algorithms to synthesize incentive sequences that enable the principal to induce the desired agent behavior.
Related Publications:
Incentive Design for Temporal Logic Objectives, Y. Savas, V. Gupta, M. Ornik, L. J. Ratliff, U. Topcu, IEEE Conference on Decision and Control, 2019 [Link]
On the Complexity of Sequential Incentive Design, Y. Savas, V. Gupta, U. Topcu, IEEE Transactions on Automatic Control, accepted [Link]
Secure Communication in the Presence of Eavesdroppers
Consider a group of robots that aim to communicate a confidential message with a base station in the presence of adversaries that eavesdrop on the wireless transmission. How can the robots ensure the security of communication under power constraints without using encryption methods that rely on the existence of a secret key shared between the robots and the base station? How would the adversaries' exact location information affect the security of communication? In this project, I propose an efficient algorithm, based on the notion of physical-layer security and convex optimization, that ensures the security of communication with probabilistic guarantees even when the exact locations of the adversaries are unknown to the robots.
Related Publications:
Physical-Layer Security via Distributed Beamforming in the Presence of Adversaries with Unknown Locations, Y. Savas, A. Hashemi, A. P. Vinod, B. M. Sadler, U. Topcu, IEEE International Conference on Acoustics, Speech, and Signal Processing, 2021 [Link]