The Role of Excitation Richness in Data-Driven Model Discovery and Reinforcement Learning Control of Nonlinear Dynamical Systems

This work investigates the role of excitation richness in data-driven model discovery and reinforcement learning control of nonlinear dynamical systems. Controlled FitzHugh–Nagumo and Van der Pol oscillators are used as benchmark systems, with dynamics identified using SINDYc and Neural Ordinary Differential Equations (Neural ODEs). Reinforcement learning controllers based on Proximal Policy Optimization (PPO) are trained on the discovered models and evaluated on the true systems. Results show that richer excitation improves identification accuracy, enhances robustness to measurement noise, and generally leads to superior downstream control performance. The study highlights excitation design as a critical component of data-driven identification and control pipelines.