Reinforcement Learning Agent for PINN Optimizer Chains

Rustam Gabdrakhmanov, Dmitry A. Gusarov, Daniel Ezhov, Alexander Hvatov

Keywords: PINN, RL, optimization chains, architectures

TL;DR: In PINNs architecture is less important than optimizer choice, best optimizers for PINNs are chains, so we find them automatically

Abstract: Physics-Informed Neural Networks (PINNs) face fundamental optimization challenges where optimizer choice dramatically outweighs architecture selection in importance. Although recent optimization chaining strategies achieve transformative performance gains, they require extensive manual tuning and domain expertise. We introduce a reinforcement-learning framework for automatic optimizer selection in PINNs, learning dynamic switching policies from training dynamics. Our approach represents optimization states through autoencoder-compressed loss landscapes, enabling the RL agent to discover connections between landscape geometry and optimal optimizer choice. This framework-agnostic method automates the construction of high-performance optimizer chains without manual intervention. Our proposed agent is evaluated on the PINNacle benchmark, where it demonstrates consistently strong results across PDE problems. We also show that the optimizer sequences automatically constructed by the agent achieve performance comparable to manually designed optimizer chains, eliminating problem-specific tuning. By transforming PINN optimization from an expert-driven process to an automated one, this work addresses a critical barrier to broader adoption of physics-informed machine learning.

Journal Opt In: No, I do not wish to participate

Submission Number: 74