Online RL-Augment: Adaptive Augmentation Policy Learning for Low-Resource Image Classification

DOI

10.2991/978-94-6239-828-3_6How to use a DOI?

Keywords

Data Augmentation; Reinforcement Learning; Deep Q-Network; Low-Resource Classification; Image Classification

Abstract

Deep learning models heavily rely on large-scale datasets to achieve robust generalization. In low-resource scenarios, models are highly susceptible to overfitting, necessitating strong regularization techniques such as data augmentation. While automated augmentation methods like AutoAugment and RandAugment have improved performance, they suffer from computationally expensive offline policy searches or rely on static magnitudes that ignore the dynamic learning state of the neural network. This paper proposes Online RL-Augment, a stable adaptive augmentation framework that integrates a Deep Q-Network (DQN) controller directly into the training loop of an image classifier. The agent dynamically adjusts the number of operations and magnitude of RandAugment at each epoch based on the network’s real-time performance. To ensure policy stability and prevent destructive transformations, we introduce an Exponential Moving Average (EMA) reward smoothing mechanism coupled with an over-augmentation penalty. Empirical evaluations on a 20% low-resource subset of the CIFAR-10 dataset using a pretrained ResNet-18 model demonstrate that the proposed method achieves an average test accuracy of 76.81% across multiple random seeds, performing competitively against static RandAugment (76.26%) without the need for manual hyperparameter tuning. Furthermore, we mitigate the inherent computational overhead of online learning by demonstrating a successful policy transfer to a 40% data subset, significantly reducing training time while maintaining high accuracy (80.12%).

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Aisha Dwi Anindita Radianto
AU  - Putri Torsia Aura Prameswary
AU  - Sabrina Ella Zafira
AU  - Wayan Firdaus Mahmudy
AU  - Novanto Yudistira
PY  - 2026
DA  - 2026/11/04
TI  - Online RL-Augment: Adaptive Augmentation Policy Learning for Low-Resource Image Classification
BT  - Proceedings of the 2026 11th International Conference on Sustainable Information Engineering and Technology (SIET 2026)
PB  - Atlantis Press
SP  - 54
EP  - 63
SN  - 2589-4900
UR  - https://doi.org/10.2991/978-94-6239-828-3_6
DO  - 10.2991/978-94-6239-828-3_6
ID  - Radianto2026
ER  -