Multi-Objective Population Based Training for Hyperparameter Optimization

Multi-Objective Population Based Training for Hyperparameter Optimization

Learn about Multi-Objective Population Based Training (MO-PBT) in this 26-minute AutoML seminar presented by Arkadiy Dushatskiy from CWI. Explore how MO-PBT extends the efficient hyperparameter optimization algorithm Population Based Training (PBT) to handle multiple conflicting objectives commonly found in real-world scenarios. Discover through experimental results how MO-PBT demonstrates superior performance compared to random search, single-objective PBT, and MO-ASHA when optimizing for multiple objectives like Precision/Recall, Accuracy/Fairness, and Accuracy/Adversarial Robustness. Access the accompanying research paper and implementation code to dive deeper into this novel approach for multi-objective hyperparameter optimization.