Co-author: Xiangrui Zeng

This work is published in SAE WCX 2022 as a techincal paper. You can find out more information here

We proposes a method to adapt backward induction, which is used to solve the vehicle speed optimal control problem for energy efficiency, to a computer with a GPU to accelerate the computation. A common application of this type of problem is to control a vehicle on a given route with surrounding vehicles, road grades, traffic signals, stop signs, speed limits, and other conditions. Several indicators can be used to determine the performance of the controller, including the energy consumption of the trip, the driving speed smoothness, and the traveling time to a given destination. Solving this optimization problem globally by backward induction is time-consuming, due to the large searching space of the vehicle’s distance, velocity, and acceleration. The proposed method converts the single thread implementation to a parallel process that runs on a consumer-level GPU. This is done by choosing the problem scale, separating independent sub-processes, and pruning the data to accommodate the GPU programming requirement. The method is tested on a simulated route with a leading vehicle, a traffic light, and speed limits. The historical behaviors of the leading vehicle are known, and they are used to predict its future behaviors in a stochastic way. Compared to the CPU-based backward induction, the proposed GPU-based version solves the given problem 15 to 30 times faster, depending on the preset granularities of variables.