Predictive Inverse Optimal Control for Linear-Quadratic-Gaussian Systems

摘要: Predictive inverse optimal control is a powerful approach for estimating the policy of an agent from observed demonstrations. Its usefulness has been established in number large-scale sequential decision settings characterized by complete state observability. However, many real decisions are made situations where not fully known to making decisions. Though extensions predictive partially observable Markov processes have developed, their applicability limited complexities inference those representations. In this work, we extend linearquadratic-Gaussian setting. We establish close connections between laws setting and probabilistic predictions under our approach. demonstrate effectiveness benefit policies that influenced partial observability on both synthetic datasets.