Kullback–Leibler control for discrete-time nonlinear systems on continuous spaces

Kullback-Leibler (KL) control enables efficient numerical methods for nonlinear optimal problems. The crucial assumption of KL is the full controllability transition distribution. However, this often violated when dynamics evolves in a continuous space. Consequently, applying to problems with spaces requires some approximation, which leads lost optimality. To avoid such paper, we reformulate problem so that it does not require unrealistic assumptions. key difference between original and reformulated former measures effort by divergence controlled uncontrolled distributions while latter replaces noise-driven transition. We show admits algorithms like one without unreasonable Specifically, associated value function can be computed using Monte Carlo method based on its path integral representation.