Learning to control switching nonlinear systems with Koopman operator regression

What the math gives to ML

The paper's transferable asset is a hybrid Koopman representation: nonlinear state evolution is approximated by a linear operator whose matrix depends on a discrete action. This creates a compact, explicitly multi-step-predictable world model without requiring a fully nonlinear recurrent predictor, while retaining action-conditioned switching structure. The most promising neural-network use is to replace or augment a learned latent dynamics model with one Koopman matrix per action and train the encoder jointly using short-horizon prediction plus long-horizon rollout losses. The resulting model can support cheap planning, stable imagination, and interpretable diagnosis of which action causes instability.