Topics in Multiagent Learning
Multi-agent systems get a rigorous treatment in this MIT course, combining game theory, optimization, and learning theory. The material starts with basic matrix games like rock-paper-scissors and builds up to imperfect information games and structured formats such as combinatorial games, polymatrix games, and stochastic games. Lessons cover equilibrium concepts, methods for computing and learning equilibria, and the computational complexity involved in finding them. The course connects these theoretical tools to recent AI breakthroughs, examining how multi-agent learning methods produced human and superhuman level agents for Go, poker, Diplomacy, and Stratego. Offered through MIT OpenCourseWare, the course includes lecture materials covering the full progression from foundational matrix games to cutting edge applications, aimed at students who already have a background in machine learning and want to understand why standard single-agent techniques like gradient descent break down when multiple learning agents interact.