Topics in Multiagent Learning

Fall 2024
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 foundataions of multi-agent systems from a combined game-theoretic, optimization and learning-theoretic perspective. We start with basic matrix games, such as rock-paper-scissors, and advance to more complex forms including imperfect information games and structured games like combinatorial games, polymatrix games, and stochastic games. Topics will cover various equilibrium concepts, aspects of equilibrium computation and learning, as well as the computational complexity of finding equilibria. Additionally, we will examine how these models and methods have driven recent breakthroughs in AI, producing human- and superhuman-level agents for well-known games such as Go, Poker, Diplomacy, and Stratego. A tentative course syllabus can be found below.

Course Info

  • Instructors: Gabriele Farina (✉️ gfarina@)
  • Time: Tuesdays & Thursdays, 11:00am - 12:30pm
  • Location: 3-333
  • TA: Zhiyuan Fan (✉️ fanzy@)
  • Office hours: see Canvas
  • Prerequisites: Discrete Mathematics and Algorithms at the advanced undergraduate level; mathematical maturity.
  • This course will use Canvas

Course Structure

The course will be lecture-based. At the end of the course there will be a few lectures of project presentations by students.

Readings will consist of a mixture of textbooks and course notes, which will be uploaded after lectures.

This course will include 2-3 homework sets and a project presentation. Projects may be done individually or in groups of 2-3 students and can be theoretical or experimental.

Grading will be as follows:

  • 50% final project
  • 50% homework sets

Schedule (subject to change)

#
Date
Topic Readings
Lecture notes
1 9/5
Introduction to the course and logistics
Syllabus Slides
Part I: Normal-form games
2 9/10
Setting and equilibria: the Nash equilibrium
Definition of normal-form games. Solution concepts and Nash equilibrium. Nash equilibrium existence theorem. Brouwer's fixed point theorem.
Nash '51 Notes html | pdf
3 9/12
Setting and equilibria: the correlated equilibrium
Topological and computational properties of the set of Nash equilibria in normal-form games. Connections with linear programming. Definition of correlated and coarse correlated equilibria; relationships with Nash equilibria.
von Stengel '24 Notes html | pdf
4 9/17
HW1
Learning in games: Foundations
Regret and hindsight rationality. Definition of regret minimization and relationships with equilibrium concepts.
Intro of BM'07 Notes html | pdf
5 9/19
Learning in games: Algorithms (part I)
General principles in the design of learning algorithms. Follow-the-leader, regret matching, multiplicative weights update, online mirror descent.
Notes html | pdf
6 9/24
Learning in games: Algorithms (part II)
Optimistic mirror descent and optimistic follow-the-regularized-leader. Accelerated computation of approximate equilibria.
SALS'15 Notes html | pdf
7 9/26
Learning in games: Bandit feedback
From multiplicative weights to Exp3. General principles. Obtaining high-probability bounds.
ZS'21 Notes html | pdf
8 10/1
Learning in games: Φ-regret minimization
Blum-Mansour construction of a swap regret minimizer. Gordon, Greenwald, and Marks general construction for Φ-regret minimizers.
GGM'08 | BM'07 Notes html | pdf
Part II: Extensive-form games
9 10/3
Foundations of extensive-form games
Complete versus imperfect information. Kuhn's theorem. Normal-form and sequence-form strategies. Similarities and differences with normal-form games.
Notes html | pdf
10 10/8
HW2
Learning in extensive-form games
No-regret algorithms for extensive-form games. Counterfactual utilities and counterfactual regret minimization (CFR).
Notes html | pdf
11 10/10
Equilibrium refinements
Sequential irrationality. Extensive-form perfect equilibria and quasi-perfect equilibrium.
Notes html | pdf
12 10/15
holiday No class
Student holiday
13 10/17
project Project ideas and brainstorming
14 10/22
project Project break
Coincides with INFORMS 2024 Annual Meeting
15 10/24
project Project break
Coincides with INFORMS 2024 Annual Meeting
16 10/29
Deep reinforcement learning for large-scale games (part I)
(Guest lecture by Samuel Sokota) Rough taxonomy of deep RL methods for games. Deep versions of fictitious play, CFR, double oracle. Game-theoretic version of PPO. Magnetic mirror descent.
Slides
17 10/31
Deep reinforcement learning for large-scale games (part II)
(Guest lecture by Samuel Sokota) Public belief states. Applications to poker, Hanabi, team games.
Slides
Part III: Other structured games
18 11/5
Combinatorial games and Kernelized MWU (Part I)
Example of combinatorial games. Kernelized multiplicative weights update algorithm.
FLLK'22 Notes html | pdf
19 11/7
Combinatorial games and Kernelized MWU (Part II)
Example of combinatorial games. Kernelized multiplicative weights update algorithm.
Notes html | pdf
20 11/12
HW3
Computation of exact equilibria
A second look at the minimax theorem. Hart and Schmeidler's proof of existence of correlated equilibria. Ellipsoid against hope algorithm.
PR'08 | JLB'15 | FP'24 Notes html | pdf
21 11/14
Markov (aka stochastic) games
Setting and taxonomy of equilibria. Stationary Markov Nash equilibria in inifinite-horizon games. Shapley's minimax theorem.
Notes html | pdf
Part IV: Complexity of equilibrium computation
22 11/19
PPAD-completeness of Nash equilibria (part I)
Sperner's lemma. The PPAD complexity class. Nash ∈ PPAD.
Notes html | pdf
23 11/21
PPAD-completeness of Nash equilibria (part II)
Arithmetic circuit SAT. PPAD-hardness of Nash equilibria.
Project break & presentations
24 11/26
project Project break
25 11/28
holiday No class
Thanksgiving
26 12/3
project Project presentations
27 12/5
project Project presentations

Below is a list of related courses at other schools.

Professor Title Year School
Farina & Daskalakis Topics in Multiagent Learning (Fall 2023 version of this course) 2023 MIT
Farina & Sandholm Computational Game Solving 2021 CMU
Aaron Roth Learning in Games (and Games in Learning) 2023 UPenn
Christian Kroer Economics, AI, and Optimization 2020 Columbia
Tuomas Sandholm Foundations of Electronic Marketplaces 2015 CMU
John P. Dickerson Applied Mechanism Design for Social Good 2018 UMD
Fei Fang Artificial Intelligence Methods for Social Good 2018 CMU
Constantinos Daskalakis Games, Decision, and Computation 2015 MIT
Yiling Chen Topics at the Interface between Computer Science and Economics 2016 Harvard
Vincent Conitzer Computational Microeconomics: Game Theory, Social Choice, and Mechanism Design 2016 Duke
Ariel Procaccia Truth, Justice, and Algorithms 2016 CMU
Tim Roughgarden Algorithmic Game Theory 2013 Stanford