MetaKoopman Enhances Dynamic System Modeling Under Shifts

Robustly modeling and forecasting nonlinear dynamics, especially when systems encounter distribution shifts, is critical for real-world decision-making. Researchers introduce MetaKoopman, a novel Bayesian meta-learning framework designed to tackle this challenge. MetaKoopman learns a Matrix Normal-Inverse Wishart (MNIW) prior over the Koopman operator, which allows it to represent complex nonlinear dynamics in a simpler, linear latent space. This framework enables closed-form Bayesian updates based on recent trajectory segments and provides a closed-form posterior predictive distribution for future states, capturing both epistemic (model uncertainty) and aleatoric (data noise) uncertainties. Evaluated on an autonomous truck and trailer system in adverse winter conditions and simulated control tasks, MetaKoopman consistently outperformed previous methods in prediction accuracy, uncertainty calibration, and robustness to various distribution shifts.