Author
Fangyu Luo
Mentors
Fatemeh Nargesian and Adam Purtee
Abstract
Multi-Objective Multi-Armed Bandit (MO-MAB) algorithms have emerged as powerful tools for solving complex decision-making problems with multiple conflicting objectives. Despite the adaptation of many MAB algorithms to multi-objective settings, the application of MAB principles within hyperparameter optimization (HPO) algorithms has not been explored extensively. This thesis delves into MO-MAB algorithms, examining their broad applications and potential to enhance HPO. A focal point of our investigation is the Hyperband algorithm [17], an HPO strategy predicated on MAB principles, specifically its integration with the fairness metric. Our analysis reveals instances where prioritizing fairness over accuracy yields superior model performance. By bridging the gap between multi-objective decision-making and machine learning model tuning, this research offers valuable insights and opportunities for further advancements in both fields.
Multi-Objective Hyperparameter Optimization with Multi-Armed Bandits