Bayesian Optimized Meta-Learning for Uncertainty-Driven Optimization

NeurIPS 2024 Workshop BDU Submission113 Authors

Keywords: Bayesian Optimization, Meta Learning, Uncertainty Optimization

TL;DR: BOML for Uncertainty-Driven Optimization

Abstract: This paper introduces a Bayesian-optimized meta-learning framework aimed at enhancing model performance in uncertain and noisy industrial environments. By integrating Bayesian Optimization with Model-Agnostic Meta-Learning (MAML), our approach dynamically fine-tunes model parameters for robust performance. The framework effectively identifies global minima despite uncertainty by utilizing Gaussian Process models with the Matérn kernel and the Maximum Probability of Improvement (MPI) acquisition function. Covariance analysis aligns training and validation losses, while L2 regularization prevents overfitting. Experimental results demonstrate the framework's ability to balance accuracy and generalization, making it suitable for diverse industrial optimization tasks.

Submission Number: 113