Vendredi 26 avril 2024

Model-based Optimization with Normalizing Flows
Sara Karami
Étudiante au doctorat, laboratoire de vision et systèmes numériques

Heure: 13h30
Local: PLT-2551

Résumé: This research aims to improve sample efficiency and accelerate convergence in Black-Box Optimization (BBO) using a model-based optimization framework. Our approach leverages a distribution-based Evolutionary Strategy (ES) improved with Normalizing Flows (NF) for better search guidance. Distribution-based ES algorithms are a common choice of search algorithm in BBO. In general, these algorithms rely on predefined distributions (e.g., Gaussian, Cauchy) to conduct queries in the search space, which introduces a potentially harmful implicit constraint for the stochastic search. To address this limitation, we propose the use of flexible distributions specifically tailored to the underlying search space, adopting NF as our generative model. Furthermore, classical BBO methods, such as CMA-ES, are susceptible to converging to local optima due to the limited capacity of their sampling distribution to capture the global structure of the objective landscape. To tackle this problem, better results are commonly achieved through multiple independent runs of the algorithm, each time varying the starting point or the method initialization and selecting the best case without using the knowledge gained in the multiple runs to improve their performance. In contrast, the adaptability of NF to the landscape in our method helps simultaneous exploration of multiple modes and prevents premature convergence to local optima. Moreover, by finetuning the neural network to update the sample distribution instead of multiple runs from scratch, our approach takes the impact of earlier samples into account, improving efficiency and outcomes.

Note: La présentation sera donnée en anglais.