New KAN framework discovers kernels in integro-differential equations

Researchers have developed a new framework for discovering memory and nonlocal kernels in integro-differential equations using constrained Kolmogorov--Arnold Networks (KANs). This approach aims to overcome limitations of existing methods that often require problem-specific derivations or restrictive assumptions. The framework utilizes two types of constrained KANs: a Monotone--Convex KAN (MC-KAN) with hard constraints and a Chebyshev-based KAN (Cheb-KAN) with soft penalties, both designed to enforce physical properties like positivity and convexity. Symbolic regression is then applied to the learned kernels to obtain interpretable closed-form representations. Experiments on various benchmarks, including a 2D nonlocal reaction-diffusion equation, showed that the hard-constrained MC-KAN was more robust than the soft-constrained Cheb-KAN when dealing with sparse and noisy data. AI

IMPACT This research could lead to more robust methods for analyzing complex systems governed by integro-differential equations, potentially impacting fields that rely on such modeling.

RANK_REASON The cluster contains a research paper detailing a new methodology for solving complex mathematical equations using neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Bernstein polynomial
• Cheb-KAN
• Chebyshev
• Kolmogorov--Arnold network
• Kolmogorov--Arnold Networks
• MC-KAN
• Salah A Faroughi

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →