AI-Enabled Online Learning and Control for Nonlinear Systems in Adversarial Environments

This project aims to develop the next generation of learning and control techniques using recent advances in Artificial Intelligence (AI) to enable a safe operation of real-world systems in adversarial environments. The vulnerability of complex societal systems to cyberattacks or digital manipulation poses a major threat to national security and public safety, where an adversary could for example manipulate data to trigger a massive blackout on power grids or could hack a few self-driving cars to cause a dangerous chain reaction and gridlock an entire traffic network. While the existing control mechanisms for dynamical systems are effective in ideal environments, they could catastrophically fail in hostile environments. The project will bring transformative change by enabling dynamical systems to operate in possibly unknown and unpredictable environments. This will be achieved by designing integrated learning and control mechanisms based on deep neural networks (DNNs) to continually learn the system’s model and possible attacks and simultaneously control the system. The intellectual merit of the project includes the development of a rich mathematical foundation that integrates AI techniques into control theory to handle unknown system dynamics and adversarial disturbances. The broader impacts of the project include advancing AI theory within unpredictable and rapidly changing environments, facilitating practical applications of AI in safety-critical systems such as emerging autonomous systems with defense applications, and strengthening national infrastructure such as power grids against advanced cyber or physical threats. The project studies dynamical systems facing unknown models subject to sophisticated, well-timed attacks that are engineered by an adversary to maximize system damage. Crucially, the system operator lacks prior knowledge of both the attack schedule and the underlying disturbance models, and yet it should control the system behavior to achieve pre-specified goals. Conventional control theory relies primarily on smooth estimators that fall short under unknown adversarial disturbances. To address this issue, this project designs a suite of non-smooth estimators with embedded DNNs to learn the model of a nonlinear system and integrate them into online control strategies. The proposed framework continually estimates the system’s model based on historical state or output data, infer possible attacks, and design optimal control actions based on the latest estimate of the model and the environment. The results of this project will be applied to the U.S. power grid, and the benefits of the proposed AI-enabled techniques over the current practices in the power industry will be qualified. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.