Séminaire en format hybride au GERAD local 4488 ou Zoom
Enabling embodied intelligence requires robots to plan according to unknown and potentially adversarial intents of interacting agents. This talk will focus on one of such scenarios where theory and methods are underdeveloped.
Specifically, we study zero-sum differential games with state constraints and one-sided information, where the informed player (Player 1) has a categorical payoff type unknown to the uninformed player (Player 2). The goal of Player 1 is to minimize his payoff without violating the constraints, while Player 2 either aims to violate the state constraints or, failing that, maximize the payoff. Examples of such games include man-to-man matchup in football and missile defense scenarios. Due to the zero-sum nature, Player 1 may need to delay information release or even manipulate Player 2's belief to take full advantage of information asymmetry, while Player 2's strategy will need to balance all possible consequences.
Existing no regret solvers (e.g., for Poker) are applicable, but are not scalable to continuous action spaces as is often the case in robotics. We will discuss efficient solvers for these games by leveraging unique structural properties of the dynamics and the game.