Model-Based Meta-reinforcement Learning for Hyperparameter Optimization

Jeroen Albrechts, Hugo Martin, Maryam Tavakol (Corresponding author)

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

Abstract

Hyperparameter Optimization (HPO) plays a significant role in enhancing the performance of machine learning models. However, as the size and complexity of (deep) neural architectures continue to increase, conducting HPO has become very expensive in terms of time and computational resources. Existing methods that automate this process still demand numerous evaluations to find the optimal hyperparameter configurations. In this paper, we present a novel approach based on model-based reinforcement learning to effectively improve sample efficiency while minimizing resource consumption. We formulate the HPO task as a Markov decision process and develop a predictive dynamics model for efficient policy optimization. Additionally, we employ the Deep Sets framework to encode the state space, which is then leveraged in meta-learning for transfer of knowledge across multiple datasets, enabling the model to quickly adapt to new datasets. Empirical studies demonstrate that our approach outperforms alternative techniques on publicly available datasets in terms of sample efficiency and accuracy.
Original languageEnglish
Title of host publicationIntelligent Data Engineering and Automated Learning – IDEAL 2024
Subtitle of host publication25th International Conference, Valencia, Spain, November 20–22, 2024, Proceedings, Part I
EditorsVicente Julian, David Camacho, Hujun Yin, Juan M. Alberola, Vitor Beires Nogueira, Paulo Novais, Antonio Tallón-Ballesteros
PublisherSpringer
Pages27-39
Number of pages13
ISBN (Electronic)978-3-031-77731-8
ISBN (Print)978-3-031-77730-1
DOIs
Publication statusPublished - 14 Nov 2024
Event25th International Conference on Intelligent Data Engineering and Automated Learning, IDEAL 2024 - Valencia, Spain
Duration: 20 Nov 202422 Nov 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume15346 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference25th International Conference on Intelligent Data Engineering and Automated Learning, IDEAL 2024
Country/TerritorySpain
CityValencia
Period20/11/2422/11/24

Fingerprint

Dive into the research topics of 'Model-Based Meta-reinforcement Learning for Hyperparameter Optimization'. Together they form a unique fingerprint.

Cite this