Validation Environment for Deep Reinforcement Learning Based Gear Shift Controllers

In conventional development processes, the control of gearshifts in automatic transmissions consists of parameter maps for open loop control and subordinate PI controllers to achieve desired target trajectories. This control approach requires tedious manual tuning by experienced engineers. Deep reinforcement learning (DRL) can be used to train neural network based controllers achieving comparable results to conventionally developed gearshifts on a transmission test bench. This article presents the validation environment (VE) with the following validation configurations (VCs) required for this comparison: Simplified transmission simulation and complete vehicle simulations, transmission test benches with simulated residual vehicle as well as test vehicles with rapid prototyping control for the transmission. The validation objectives, the test bench and common interfaces are discussed. Furthermore, several Key Performance Indicators (KPIs) as evaluation criteria for gear shift criteria are presented and checked for suitability. Two methods for calculating the correspondence between the VCs are presented, including the dynamic time warping (DTW) method. ... mehrThe validity of the VCs and KPIs are shown using preliminary results.