CMPUT 676: Optimization and Decision-Making under Uncertainty
Fall 2024
Instructor
: Xiaoqi Tan (xiaoqi.tan$\textsf{@}$ualberta.ca)
Location & Time
: GSB 7-11, MW 12:00 PM – 1:20 PM
Office hour
: After class or by appointment
Course Overview
Many real-world problems involve making decisions over time in the presence of various forms of uncertainty. These challenges arise in fields such as Internet advertising, energy sustainability, transportation, financial trading, healthcare, and a wide range of problems in artificial intelligence and machine learning. Although these specific decision problems may initially appear different in various application scenarios, the models and algorithms needed to address them are often similar.
In this research-oriented course, we will review recent developments and discuss open directions in the general field of decision-making under uncertainty through several modern optimization lenses. We will begin with a brief introduction to convex optimization, covering basic concepts of convex sets and convex functions, canonical convex problems, Lagrange multipliers, duality theory, KKT optimality conditions, and standard convex optimization algorithms. Following this, we will explore models and algorithms designed to handle different forms of uncertainty in optimization and decision-making. Major topics include: i) online optimization and competitive analysis; ii) stochastic optimization, learning, and approximation; and iii) algorithmic game theory and mechanism design. These topics are highly interdisciplinary, with strong connections to theoretical computer science, artificial intelligence, machine learning, economics, operations research, statistics, and control. While the course is theoretical in nature, most problems discussed will be motivated and illustrated by practical examples.
Notes: We welcome high level undergrads to audit or take the course (subject to administrative approval).
Course Project
The main assignment of this course is to complete a course project which includes a proposal (20%), an in-class presentation (30%), and a final report (50%).
Proposal (20%): You need to select 1-2 papers from this
list
and prepare a one-page review (covering the main idea, methodologies, and key results of the selected papers) and a separate one-page proposal (containing potentially multiple ideas, such as improvements, alternative approaches, or new discoveries). Your paper review and proposal will be evaluated using a scoring grid based on the following criteria: i) depth, rigor, and thoroughness of the review, ii) clarity and merit of the proposal, and iii) coherence between the review and the proposal. Please note that the one-page proposal is proportionally more important, which is why the task is titled “Proposal” rather than “Paper Review.”Presentation (30%): You will give an in-class presentation based on your paper review and proposal. Your presentation will be evaluated using a scoring grid based on the following criteria: i) clarity of the content (e.g., structure of the slides), ii) clarity of the presentation, and iii) effectiveness in addressing questions. These three criteria are equally important in evaluating your presentation.
Final Report (50%): You are expected to complete this course with an 8-10 page final report. As a research-oriented graduate course, it is up to you to decide the focus of your final report—whether it be a proof of a new theorem, a new implementation, or a new survey—but it must be related to your proposal. Your final report will be evaluated using a scoring grid based on the following criteria: i) novelty, ii) consistency with the proposal and presentation, and iii) relevance to this course. Please note that novelty does not necessarily mean novel research results. While you will receive bonus marks for making novel research discoveries relevant to this course, they are not required to receive full marks on the project.
Course Schedule and Readings
Date | Topics | References and Readings |
---|---|---|
Sept. 4 | Lec 1: Overview Course topics; Logistics |
L1-slides |
Convex Optimization (A brief intro) |
||
Sept. 9 | Lec 2: Concepts Convex sets Convex functions Convex problems |
L2-slides BV Book: Ch1-Ch4 |
Sept. 11 | Lec 3: Theory Lagrange multipliers Duality theory Optimality conditions |
L3-slides BV Book: Ch5 DB Book |
Sept. 16 | Lec 4: Algorithms Gradient descent Newton’s methods Barrier methods Dual ascent |
L4-slides BV Book: Ch9-Ch11 |
Online Optimization |
||
Sept. 18 | Lec 5: Online Algorithms (Basics) Ski rental problem Deterministic vs Randomized Yao’s principle |
L5-slides BP book: Ch1-Ch3 Ski rental |
Sept. 23 | Lec 6: Online Primal-Dual (OPD) Time series search $k$-max and $k$-min search One-way trading The OPD approach (1st look) |
L6-slides Online search and one-way trading $k$-max and $k$-min search Online selection with convex costs |
Sept. 25 | Lec 7: Online Knapsack Threshold-based algorithms The OPD approach (2nd look) |
L7-slides, notes Online knapsack and variants Online knapsack with convex costs Generized one-way trading; multiple knapsack Knapsack and advice complexity |
Sept. 30 | ||
Oct. 2 | Lec 8: Online Matching RANKING algorithm The OPD approach (3rd look) |
L8-slides AM Survey, OPD Survey Readings on RANKING: KVV; An elegant proof; Economic-based analysis; Randomized OPD analysis |
Oct. 7 | Lec 9: Online Convex Optimization (OCO) The Experts problem OCO: intro + algorithms + variants Competitive ratio vs. Regret (1st look) |
L9-slides EH Book, CBL Book |
Oct. 9 | Lec 10: Online Linear Programming (OLP) OLP: intro + variants Adversarial inputs vs. Random inputs Competitive ratio vs. Regret (2nd look) |
L10-slides Random-order models Online LP under ROM Online LP under IID |
Oct. 14 | ||
Oct. 16 | Lec 11: Beyond Worst-Case Analysis (WCA) Beyond WCA: why + how Learning-augmented algorithms Data-driven algorithms |
L11-slides Beyond WCA Algorithms with predictions Data-driven algorithm design |
Stochastic Optimization and Learning |
||
Oct. 21 | Lec 12: Markov Decision Process (MDP) MDPs: A brief intro Algorithms: Value/policy iteration; LP approach Example: energy storage control |
L12-slides Courses: CMPUT 609, CMPUT 653 |
Oct. 23 | Lec 13: Multi-Armed Bandits (MABs) MABs: intro + variants Stochatic bandits; Algorithms |
L13-slides LS Book, BCB Book, Slivkins Book |
Oct. 28 | Lec 14: Stochastic Approximation The Robbins-Monro algorithm Stochastic gradient descent (SGD); Adaptive SGD Summary: Online vs. Stochastic |
L14-slides |
Algorithmic Game Theory; Mechanism Design |
||
Oct. 30 | Lec 15: Game Theory Pure-strategy Nash equilibrium Mixed-strategy Nash equilibrium |
L15-slides |
Nov. 4 | Lec 16: Mechanism Design (MD) Key concepts of MD Auctions; Vickrey-Clarke-Groves mechanisms Optimal auctions |
L16-slides MD reading: Econ, CS Auctions reading: Survey |
Nov. 6 | Lec 17: Online Mechanism Design (OMD) OMD: intro Online auctions Posted price mechanisms Prophet inequalities Course summary |
L17-slides |
Nov. 11 | ||
Nov. 13 | Proposal Due (eClass link) | |
Project Presentations |
||
Nov. 18 | ||
Nov. 20 | ||
Nov. 25 | ||
Nov. 27 | ||
Dec. 2 | ||
Dec. 4 | ||
Dec. 9 |