Prof. Mushuang Liu, University of Missouri

Title: Game-Theoretic approaches for Autonomous Driving : Challenges and Possible Solutions

Abstract: Fully autonomous vehicles (AVs) are envisioned to bring substantial societal benefits, including improved road safety, reduced traffic congestion, and enhanced mobility for the disabled and elderly. Despite the significant advantages, major technical challenges still remain to be addressed before AVs can routinely drive on public roads. One of the most critical challenges lies in the design of the AV decision-making algorithms. To enable interactive AVs with human-like reasoning skills and behavior, game-theoretic approaches have been explored in recent developments. In game theory, agents are assumed to pursue self-interests while considering the interactions with others. Such a reasoning process is often considered consistent with humans’ (and thus human drivers’) reasoning and decision-making, facilitating interaction-aware AV decision-making algorithm design. Along this line of research, both Nash games and Stackelberg games have been explored in the context of autonomous driving. However, these game-theoretic approaches often suffer from at least one of the two major challenges: Game complexity and Incomplete information. Game complexity refers to the difficulties of solving a multi-player game. These difficulties include but are not limited to (a) solution existence, (b) algorithm convergence, and (c) scalability. Incomplete information refers to the ego vehicle, i.e., the AV, lack of global knowledge of other traffic agents’ objective functions or cost functions, considering the variability of human driving habits or styles. In this talk, we will provide possible solutions to addressing these challenges. The latest techniques developed to advance the practicality and robustness of game-theoretic approaches in autonomous driving will be discussed.