Topics in Multiagent Learning

While machine learning techniques have had significant success in single-agent settings, an increasingly large body of literature has been studying settings involving several learning agents with different objectives. In these settings, standard training methods, such as gradient descent, are less successful and the simultaneous learning of the agents commonly leads to nonstationary and even chaotic system dynamics.

Motivated by these challenges, this course presents the foundations of multi-agent systems from a combined game-theoretic, optimization and learning-theoretic perspective, building from matrix games (such as rock-paper-scissors) to stochastic games, imperfect information games, and games with non-concave utilities. We will present manifestations of these models in machine learning applications, from solving Go to multi-agent reinforcement learning, adversarial learning and broader multi-agent deep learning applications. We will discuss aspects of equilibrium computation and learning as well as the computational complexity of equilibria. We will also discuss how the different models and methods have allowed several recent breakthroughs in AI, including human- and superhuman-level agents for established games such as Go, Poker, Diplomacy, and Stratego. A tentative course syllabus can be found below.