Researchers have developed a novel zeroth-order deep learning method to tackle high-dimensional partial differential equations (PDEs) with unknown coefficients, a common challenge in scientific machine learning and continuous-time reinforcement learning. This new approach bypasses the instability and derivative errors associated with repeated automatic differentiation in high dimensions by using only function evaluations. The method employs perturbed Monte Carlo trajectories to estimate derivatives, enabling a fully model-free approach that generates targets for gradient and Hessian networks. A statistical analysis demonstrates the method's effectiveness, providing error bounds and characterizing sample complexity in weighted Sobolev spaces, with numerical experiments showing competitive performance in moderate and high dimensions. AI
IMPACT Introduces a novel, model-free deep learning technique for solving challenging PDEs, potentially advancing scientific computing and reinforcement learning applications.
RANK_REASON Academic paper detailing a new methodology for solving complex mathematical problems. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Continuous-Time Reinforcement Learning
- Scientific Machine Learning
- Zeroth-Order Deep Learning Method
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
