TY - GEN
T1 - Model-Based Meta-reinforcement Learning for Hyperparameter Optimization
AU - Albrechts, Jeroen
AU - Martin, Hugo
AU - Tavakol, Maryam
PY - 2024/11/14
Y1 - 2024/11/14
N2 - 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.
AB - 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.
UR - http://www.scopus.com/inward/record.url?scp=85210473048&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-77731-8_3
DO - 10.1007/978-3-031-77731-8_3
M3 - Conference contribution
AN - SCOPUS:85210473048
SN - 978-3-031-77730-1
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 27
EP - 39
BT - Intelligent Data Engineering and Automated Learning – IDEAL 2024
A2 - Julian, Vicente
A2 - Camacho, David
A2 - Yin, Hujun
A2 - Alberola, Juan M.
A2 - Nogueira, Vitor Beires
A2 - Novais, Paulo
A2 - Tallón-Ballesteros, Antonio
PB - Springer
T2 - 25th International Conference on Intelligent Data Engineering and Automated Learning, IDEAL 2024
Y2 - 20 November 2024 through 22 November 2024
ER -