Modeling Human Strategic Behavior
(CPSC 532L/530L, Term 2, 2021-22)

 

Overview

Course Description: This course will focus on models for predicting human strategic behavior. At the beginning of the course, groups of students will identify a strategic setting (e.g., from their own research areas) which they will study throughout the course. We will learn about how to analyze strategic settings from the perspective both of classical game theory and of machine learning, and students will apply such analysis to their chosen domains. The last part of the course will survey advanced work in modeling strategic settings from both machine learning and behavioral game theory.

 

 

Co-Instructor: The course will be remotely co-taught by Prof. James R. Wright of the University of Alberta. Classes will be held over Zoom to permit full interaction with U of A students. Students will be free to do group work with other UBC and/or U of A students, as they prefer.


Meeting Times: Tuesday, Thursday, 2:00 PM - 3:30 PM

First Class: Thursday, January 11, 2022

Location: ICICS/CS Building, room 246

UBC Web Pages: 532L; 532L waiting list; 530L; 530L waiting list

Instructor: Kevin Leyton-Brown

Instructor's Office Location: ICCS X565

Instructor's Office Hours: Tuesdays and Thursdays 3:30 - 4:00 PM, or by appointment

 

Prerequisites: The course has no formal prerequisites. As a graduate topics course, it will survey current research literature and expect students to be able to read, summarize, and form critical opinions of this material. Students may find it useful to have background in machine learning and in microeconomics and game theory; however, I expect that many students will not have all of this background. (Particularly, I recognize that most CS students may not have previous exposure to economics.) Data analysis and basic coding will be required. Additionally, an ability to speak, read and write fluently in English is essential for success in the class.

 

Course number: The course is cross-listed as CPSC 532L (Topics in Artificial Intelligence, part of the department's "Computational Intelligence" stream) and CPSC 530L (Topics in Information Processing, part of the department's "Interdisciplinary Studies" stream). Students are free to enroll in whichever course better suits their needs.

 

Equity, Inclusion and Wellness: Please see the CS Department's resources on this topic.

 

Academic Honesty: Plagiarism is a serious offence (see the CS Department's statement) and will be dealt with harshly.  I consider plagiarism to be the unattributed use of an external source (e.g., another student, code or text from a web site, a book) in work for which a student takes credit, or the inappropriate use of an external source whether or not attribution is made. The seriousness of the offence depends on the extent to which the student relied upon the external source.  You must cite all external sources that you use, and write in your own words. Any text that you take verbatim from another source must be in quotation marks and followed by a citation.


Syllabus

The course will consist of four major units.

 

Unit Deliverables

1. Modeling Strategic Situations. We will begin by exploring what is meant by a "strategic situation": roughly, an environment in which multiple self-interested actors interact, and in which their satisfaction with the resulting state of the world is based on the decisions that both they and the other actor(s) chose. We will consider a variety of models of such interactions (simultaneous moves; sequential moves, with both perfect and imperfect information; Bayesian uncertainty; infinite repetitions of any of the above). We will also consider what makes a setting inappropriate for consideration as strategic: e.g., decision-theoretic settings exhibiting weak or no coupling between agents' payoffs.

 

Students will form small groups and identify a strategic domain that they will study together for the remainder of this course.

We strongly encourage students to choose domains related to their existing research interests and/or expertise, and to form groups with students having similar interests.

2. Game Theoretic Solution Concepts. This unit will survey the classical game theoretic question: "How should strategic agents behave?" We will learn various game theoretic answers to this question, such as best response, Nash equilibrium, dominant and dominated strategies, minimax strategies, and minimax regret strategies. We will learn how to apply these answers in a variety of game formalisms, including normal-form, extensive-form games with both perfect and imperfect information, Bayesian games, and repeated games.

 

Students will identify various games modeling aspects of their target domains and making use of different game formalisms. For each, they will identify game theoretic solution concepts, and will argue which is most appropriate for understanding the domain.

3. Modeling Strategic Behavior as a Machine Learning Problem. Unfortunately, humans often behave differently from the way game theory predicts. In this unit, we will consider an alternate approach to modeling human behavior in strategic settings, leveraging techniques from machine learning. We will consider a variety of candidate model families, including the Level-k, Cognitive Hierarchy, Quantal Response, and Quantal Level-k models from the behavioral game theory literature, and will also learn why traditional machine learning model families are often inappropriate for use in this setting. We will learn how to design experiments, obtain training data from human play of games, fit models to this training data, and detect and avoid overfitting these models.

 

Students will play each others' games from the previous unit and gather training data. They will use this data to fit different model families, and will argue which is most appropriate to their domain.

4. Advanced Topics. The last unit will survey cutting-edge research on modeling human behavior by combining methods from behavioral game theory and machine learning. Some possible topics to be covered in this unit include no-regret learning, using deep learning to play large sequential, zero-sum games such as go or poker, computational approaches for reasoning about large general-sum games, refinements to the BGT model families described above such as more sophisticated level-zero models, and behavioral models of finance.

At the end of the course, student groups will hand in a hypothetical thesis proposal putting forward a research program for modeling human strategic behavior in their chosen domain. Like a real thesis proposal, this will explain why the problem is important, survey related literature, present initial results, and describe promising avenues for further exploration.

 

Deliverables & Grading

Evaluation will consist of the following elements: