A trust-region formulation is proposed for sampling-based model predictive control algorithms such as MPPI, constraining updates of the proposal distribution via a KL divergence bound and, optionally, an entropy lower bound. This replaces heuristic tuning of temperature or momentum with values optimal with respect to the underlying Lagrangian.
Principled hyperparameter tuning improves predictability and reproducibility of sampling-based controllers in complex system management.
Translation and summary are machine-generated from the source. The full article is not republished; its rights belong to the source.
