Information-Theoretic Adaptive Cooling for Deterministic MPPI via Entropy Feedback

What the math gives to ML

The paper offers a constructive feedback mechanism for controlling temperature in deterministic MPPI: use the Shannon entropy of normalized importance weights to detect whether sampling remains exploratory or has prematurely collapsed. Its transferable asset is an online entropy controller with a critical threshold that changes the cooling rate before effective sample size becomes too small. A direct neural-network transfer is a derivative-free population optimizer for model parameters, prompts, or architecture variables, where candidate losses generate importance weights and entropy controls the candidate distribution. The key falsifiable signature is a kink in the temperature-decay rate and a corresponding transition in effective sample size when normalized entropy crosses the chosen threshold.