Adaptive Data Harvesting for Efficient Neural Network Learning with Universal Constraints

Researchers Siteng Kang and Xinhua Zhang have proposed a novel method titled 'Adaptive Data Harvesting' to enhance the training efficiency of neural networks subject to universal constraints over continuous domains. This approach specifically addresses challenges in Lyapunov Neural Networks (Lyapunov NNs) and Physics-Informed Neural Networks (PINNs), where analytical solutions are often unavailable or overly restrictive. Traditional sample-based methods typically rely on fixed heuristics or handcrafted rules, which can lead to suboptimal convergence speed, stability, and solution quality. In contrast, this new method employs reinforcement learning to dynamically and iteratively adjust data samples based on the model's evolving performance. By learning from data and experience, the system optimizes sample selection in real-time. The authors validated their approach on both Lyapunov NNs and PINNs, demonstrating significant improvements in empirical constraint satisfaction and overall training efficiency. This research highlights the broader applicability of adaptive input selection in domains where effective training is critical, offering a robust alternative to static sampling strategies in machine learning applications involving complex physical or mathematical constraints.