Adaptive Active Learning for Regression via Reinforcement Learning
New reinforcement learning agent dynamically selects which data to label, outperforming 18 benchmark datasets.
A team of researchers has published a new paper, "Adaptive Active Learning for Regression via Reinforcement Learning," introducing Weighted improved Greedy Sampling (WiGS). This method tackles a core challenge in machine learning: the high cost of labeling data for regression tasks. It improves upon the established Improved Greedy Sampling (iGS) technique by replacing its static, multiplicative rule for selecting data points with a dynamic, additive criterion. The key innovation is framing the weight selection for this criterion as a reinforcement learning problem, allowing an AI agent to learn and adapt the optimal balance between exploring diverse data and investigating uncertain predictions throughout the training process.
Experiments demonstrated WiGS's superior performance across 18 diverse benchmark datasets and a synthetic environment. The method consistently outperformed iGS and other baselines in both final model accuracy and labeling efficiency, meaning it achieved better results with fewer expensive labeled examples. The paper highlights that WiGS is particularly effective in domains with irregular data density, where previous methods could ignore high-error samples clustered in dense regions. The full research, submitted to UAI 2026, includes 8 main pages with 4 figures and an extensive appendix, and the codebase has been made publicly available for further development and application.
- WiGS replaces a static sampling rule with a reinforcement learning agent that dynamically adjusts data selection strategy.
- Tested on 18 benchmark datasets, it achieved higher accuracy with greater labeling efficiency than previous methods.
- It excels with irregular data distributions, solving a flaw in prior methods that ignored errors in dense data clusters.
Why It Matters
Dramatically reduces the time and cost of preparing training data for real-world predictive models in science and industry.