Effective black-box optimization is a widespread challenge in many real-world applications, including recommender systems, computational biology, drug discovery, and hyperparameter tuning of large workflows. Bayesian optimization has been a state-of-the-art technique, and among the most popular approaches that empower these modern applications.
This course includes a tutorial (first three weeks) on Bayesian optimization, covering the core concepts and fundamental algorithms, followed by discussions of research papers exhibiting the key practical challenges and recent advancements. Students will read and discuss recent research papers, lead student lectures (from week 4-9), as well as conduct a research project in 1-2 student groups.
Lectures on Tu/Th 2:00pm-3:20pm in RY 277
Office hour (by appointment): Tu/Th 3:30pm-4:00pm in JCL 317
First lecture: 1/3/2023
Announcements on Canvas
Discussion via Ed Discussion
30%: Paper reviews and lecture presentation
70%: Research project (Report + poster presentation)
CMSC 25300/CMSC 25400/TTIC 31020 or equivalent, or permission from instructor
Week 1: Introduction
Week 2: Practical challenges: Scalability & kernel learning
Week 3: Advanced BO methods and open problems.
Week 4: Noisy & indirect observations
Week 5: Complex objectives
Week 6: Complex constraints
Week 7: Complex dynamics
Week 8: Massive search space
Week 9: Delayed feedback
Bayesian Optimization, by Roman Garnett, Cambridge University Press, January 2023.
Active Learning, (accessible via UChicago IPs), by Burr Settles, Morgan & Claypool Publishers, 2012