CAREER: Game-theoretic Online Planning in Partially Observable Domains

Autonomous systems enabled by artificial intelligence (AI) have the potential to work alongside humans to accomplish previously impossible goals in science, transportation, healthcare, manufacturing, and defense. One critical challenge is designing these systems so that they interact with humans and other AI systems in a predictable and safe way that is beneficial to humans and robust to attacks from bad actors. If this challenge is addressed properly, self driving cars could deliver their passengers quickly through traffic without compromising the safety of other road users, teams of drones could monitor a thunderstorm with limited communication bandwidth to predict tornadoes, and networks of telescopes could monitor low-earth orbit to prevent satellite collisions and protect critical space systems from sabotage. Game theory is a mathematical framework that describes interaction between intelligent human or AI decision-makers called "agents" and prescribes the best strategies for accomplishing goals through interaction. Game theory can model agents with goals that are aligned, opposed, or somewhere in between. It can reveal complex and surprising interaction patterns including ways that agents with disparate goals can cooperate for everyone's benefit, strategies to prevent deception or exploitation by a disruptive agent, and interaction that might result in danger for all agents. In this project, the team will develop decision-making algorithms based on game theory that allow autonomous agents to interact and achieve their goals safely and efficiently even when there is uncertainty about the environment or other agents. In order to disseminate findings rapidly so that they can be applied to real-world problems, the team will develop an open-source programming toolbox, create new university-level course materials, and work with high school teachers to begin developing a diverse next generation of AI engineers. Most recent AI advances, including generative AI systems, are created using an offline training process, meaning that they are trained with vast amounts of data before interacting with the environment or making a decision. This project will focus on a complementary approach called online planning, where reasoning is carried out at the time of interaction with the environment. Online planning has strengths in explainability and ease of composing multiple models, and it can be used to speed up offline learning. This research will focus specifically on developing algorithms for partially observable stochastic games (POSGs), building on recent advances in single agent optimization in partially observable environments. The algorithms will be designed to handle large continuous state and observation spaces that real-world cyber-physical systems act in. The first phase will tackle the cooperative case, where all agents have the same goal, but must still make decentralized decisions with incomplete information, acting in a predictable way to achieve common objectives. The second phase will shift to zero-sum POSGs, where agents' goals are in direct opposition. In this setting, the agents seek to protect information from each other and avoid exploitation by acting in a less-predictable way. Finally, the third phase will focus on general-sum POSGs where goals are neither perfectly aligned or opposed. Here, a key challenge is to use conventions that allow for coordination and negotiation. The result of this research will be a palette of easy-to-use and trustworthy algorithms that engineers can apply to many real-world systems that involve interaction and uncertainty. 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.