Optimizing Software Parameter Tuning with Software-in-the-Loop - A Comparative Study of Black-Box Optimization Methods for the Calibration of Transmission Software

Abstract

Modern vehicle software systems are complex and contain numerous tunable input parameters that influence the behaviour of the vehicle. Configuring and testing these by hand is time-consuming and expensive. To streamline and automate this process, Software-in-the-Loop (SiL) is used to iteratively run a simulation of the vehicle and its surrounding environment whilst updating the vehicle configuration. This thesis investigates various optimization methods for automatically tuning software parameters in the software system in a Battery Electric Vehicle (BEV) using SiL. The objective is to minimize the time between a requested signal and the actual signal. Due to its simulation-based nature, the objective function is gradient-free and expensive and is therefore treated as a Black-Box Optimization (BBO) problem. Investigated methods include Bayesian Optimization (BO), Particle Swarm Optimization (PSO), Covariance Matrix Adaption Evolution Strategy (CMA-ES), and Particle Swarm Optimization Bayesian Optimization (PSOBO), as well as baseline methods for comparison. The results indicate that advanced optimization algorithms consistently achieve lower objective values than the baseline methods. CMA-ES demonstrates the best performance across all test cases. Future work includes investigating the transferability of the optimized parameters from the simulation to a physical vehicle.