カルバック・ライブラ情報量の非対称性に着目したサンプリングベースモデル予測制御

抄録

Model Predictive Control (MPC) is one of the effective control methods for complex systems such as automatic driving and robotics. As one of the MPC solvers, the cross-entropy method (CEM) is well known as the most flexible and general method. Although CEM can be applied to most systems, it requires a sufficient (theoretically infinite) number of samples and updates for convergence, resulting in extremely high computational cost. Therefore, we focus on the asymmetry of the Kullback-Leibler divergence used in the minimization problem of CEM, and propose a new algorithm for CEM by redefining its minimization problem, so-called risk aversion CEM (RA-CEM). RA-CEM allows the function that can be regarded as a weight for the sampled trajectory to take negative values, so that even with a small iteration, the algorithm actively avoids trajectories with poor performance and prioritizes convergence to trajectories with good performance. In a highway driving simulation, RA-CEM improved the success rate from the standard CEM.