
Seminar: Gittins Indices for Cost-aware and Freeze-thaw Bayesian Optimization
February 18 @ 2:00 pm - 3:00 pm
Presented by Qian Xie, Cornell
Hyperparameter optimization is crucial in real-world applications such as machine learning model training, robotics control, material design, and plasma physics. In transportation, hyperparameter optimization plays a significant role in applications like traffic flow prediction, dynamic pricing, route planning, and public transportation scheduling, where complex models need to be fine-tuned to achieve optimal performance. These scenarios are often modeled as black-box functions, which take hyperparameters as inputs and output performance metrics. Bayesian optimization is a powerful framework for efficiently optimizing such black-box functions, especially when evaluations are time-consuming or expensive. However, practical factors such as varying function evaluation costs and observable partial feedback during function evaluation remain under-explored in this framework. My research leverages Gittins indices, which are inherently cost-aware and feed
In the first half of my talk, I will present my published work, which adapts Gittins indices into a cost-aware acquisition function class for Bayesian optimization, demonstrating competitive empirical performance, particularly in medium-to-high dimensions. In the second half, I will discuss my ongoing work on developing Gittins indices for freeze-thaw Bayesian optimization involving decisions on early stopping and switching of hyperparameter tests based on partial feedback.