In AI and beyond, systems of multiple agents are naturally modeled using game theory. From game theory, we know that sometimes, when each agent pursues its own objectives, the outcome may be one that is bad for all agents (e.g., the Prisoner's Dilemma). Learning algorithms can indeed converge to such bad equilibria. What can be done to prevent such bad outcomes, and how should we think about designing agents in such contexts? In this course, we will approach this question from a variety of angles, ranging from traditional approaches in game theory to novel ones that fit AI better than humans.
PREREQUISITES:
There is no formal prerequisite for the course, but we do expect students to be mathematically well prepared and ready to undertake a significant course project. For students just looking to gain general background in AI, 15-780 is better suited.
TEXT:
Materials will be made available on the course website. A text that provides general background in game theory is Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations by Shoham and Leyton-Brown, but this text covers only some of the topics of the course.
METHOD OF EVALUATION:
Grading will be based on class participation (10%), homework assignments (20%), a midterm exam (20%), and a class project (50%).
TOPICS LIKELY TO BE COVERED:
Game theory: representations, solution concepts, algorithms
Cooperation in repeated games and stochastic games, folk theorems
Commitment
Program equilibrium
Correlated equilibrium, mediated equilibrium
Team games
Mechanism design
Automated mechanism design
Market design and algorithms for running such markets, e.g., kidney exchange
Learning in games, equilibrium selection
Evaluating agents, also in non-zero-sum games
Agent design: identities, preferences, beliefs
Imperfect recall, belief formation, variants of decision theory