Inverse Reinforcement Learning for Identification in Linear-Quadratic Dynamic Games

Abstract

The theory of dynamic games has received considerable attention in a wide range of fields. While great effort has been made to develop new algorithms for finding Nash equilibria in dynamic games, the identification of cost functions has received little attention. We present an identification algorithm for linear quadratic dynamic games, a framework which can be applied in the field of shared control between a human and an automatic controller. In this application, the cost function describing human behavior is identified, taking into account the influence of the automation. Furthermore, we consider that human movement underlies certain variability by using a probabilistic Inverse Reinforcement Learning approach. As identification is performed in a single optimization step, the proposed method is suited for real-time applications. A simulation example shows that the algorithm successfully identifies the cost function of the first player which—in combination with the second player—reproduces the observed system output.