CSCI699: Topics in Learning and Game Theory (Fall 2017)
- Lecture time: Monday 3:30pm - 6:50pm
- Lecture place: KAP 145
- Instructors: Ilias Diakonikolas, Shaddin Dughmi
- Emails: firstname.lastname@example.org, email@example.com
- Office: SAL 220 (Ilias), SAL 234 (Shaddin)
- Office Hours: By Appointment
- TA: Li Han
- Email: firstname.lastname@example.org
- Office Hours: Monday 2pm-3pm and Thursday 2pm-3pm, in SAL 246
- Course Homepage: http://www-bcf.usc.edu/~shaddin/cs699fa17/index.html
- Nov 19: Homework 3 is out. It will be due on Friday Dec 1. Those of you who have scribed twice only need to solve 40 of the 50 points for full credit.
- Oct 27: Homework 2 is out. It will be due on Friday Nov 10.
- Sep 20: Homework 1 is out. It will be due on Wednesday October 4.
- Sep 5: Scribing template is here. Scribe notes are due 1 week after lecture.
- Sep 5: Scribe notes for the first 2 lectures have been posted.
- Aug 29: Course Mailing list has been created. You should have received an invite. If not, email Li and he can add you.
- Aug 17: Course website is up!
Schedule by Week
- Weeks 1-2: Basics of Learning Theory
- Weeks 3-4: Basics of Game Theory and Mechanism Design
- Weeks 5-6: No-Regret Learning and Equilibria
- Weeks 7-8: Mechanism Design from Samples
- Week 9: Mechanism Design from Revealed Preferences
- Week 10-15: Various topics on collecting good data from self-interested agents for ML/other applications. Based on time available and popular demand, we will cover a selection of the following topics:
- Eliciting distributions using proper scoring rules
- Peer prediction
- Bayesian truth serum
- Prediction markets
- Incentivizing exploration in online learning
- Buying private data
This course will examine recent research trends at the interface of learning and game theory. Topics will include online learning and its connection to game equilibria, the design of auctions and mechanisms from data, computational aspects of econometrics, learning from strategic data sources, etc. This class will be targeted at PhD students. Mathematical maturity, as well as research experience in computer science and/or data science is strongly recommended.
- Lectures 1 and 2 (August 21 + 28) Introduction to machine learning theory: PAC Learning. Empirical Risk Minimization. Uniform Convergence. VC Dimension. Radermacher Complexity. Unsupervised Learning.
- Lecture 3 (Sept 11) Introduction to Game theory
- Lecture 4 (Sept 18) Introduction to Mechanism Design
- Lecture 5 (Sept 25) Online learning and the Multiplicative Weights Algorithm
- Lecture 6 (Oct 2 + first half of Oct 9) Online learning and convergence to equilibria: The minimax theorem, coarse correlated and correlated equilibrium, and reduction from swap regret to external regret.
- Scribe notes
- Reading: Chapter 4.4 - 4.5 of the AGT book.
- Additional reading: Tim Roughgarden's lecture notes here and here
- Lectures 7+8 (Oct 9 + Oct 23 + first half of Oct 30) Mechanism Design from Samples
- Lecture 9 (Oct 30) Eliciting distributions using proper scoring rules.
- Lectures 10+11 (Nov 6 + Nov 13) Crowdsourcing Information: Peer Prediction and the Bayesian Truth Serum.
- Prerequisite Courses: No formal prerequisites.
- Recommended Preparation: Exposure to machine learning, game theory, and theoretical computer science, with research experience in at least one, is recommended. Mathematical maturity and an interest in exploring research questions in the area are expected.
Requirements and Grading
- There will be 3-4 homework assignments worth 40% of the grade. Collaboration and discussion among students is allowed for the homeworks, even encouraged, though students must write up their solutions independently.
- Scribing duties will be divided among students in the class, and will count for 10%. Scribing template is here. Notes are due 1 week after lecture.
- Class participation will count for 10%.
- A research project, due at the end of the semester, will count for 40%. Students will have to choose a related research topic, read several papers in that area, and write a survey of the area. More details and ideas will follow in due course.
Late Homework Policy: Students will be allowed 4 late days for homework, to be used in integer amounts and distributed as the student sees fit. Additional late days will each result in a deduction of 10% of the grade of the corresponding assignment.
This is a topics course, and therefore we will refer to many sources including books and papers; in the vast majority of cases those sources will be available online, and will be linked on the course homepage.