Acceleration Control Design of HEVs with Comfortability Evaluation based on IRL

Abstract For passenger cars, comfortability is an important issue to consider in powertrain control. However, it is not easy to take comfortability into account when designing a powertrain control strategy because of its subjective nature and difficulty in being quantified. This paper presents a solution by using the inverse reinforcement learning (IRL) method. An acceleration scenario of a hybrid electric powertrain is considered to show this design approach. With a sample acceleration profile scored by an expert evaluating module, a reward function is obtained by training an extreme learning machine. Using the analytical representation of comfortability, an estimation of distribution algorithm (EDA) is used to seek the optimal acceleration reference. Given the reference acceleration signal, the control law for the electric motors that provide power assistance during the acceleration phase is obtained by solving a optimal tracking control problem. A numerical example is shown to evaluate the design approach.