Modern powertrain is evolving into electrified, networked, software-intensive systems with a large number of degrees of freedom. This results in an enormous increase in system complexity with shorter development cycles and thus an exponentially increasing effort for the development, validation and calibration of software functions.
Due to the increasing number of variants and limited development capacity, conventional development processes with human-created software lead to suboptimal solutions. The highly nonlinear relationships between single components, manipulated variables, parameters, and perturbations are difficult to handle even for experts, and the functions developed are usually not applicable to similar problems without considerable additional effort. With the limited availability of skilled personnel, this leads to unexploited potential for efficiency improvement that can be leveraged by machine learning.
The resolution of the conflicting goals of efficiency, emissions, fulfillment of the driver's desire and other application-specific target variables is addressed at the Institute of Mechatronics for Mobile Drives by reinforcement learning, among other methods. This machine learning method is based on the derivation of optimal strategies to maximize a given reward function based on experience. Although the fundamental potential of this methodology has already been demonstrated in numerous applications, reinforcement learning in the context of powertrain software has been used to an extremely limited extent, and represents a clear novelty.
In order to create application-oriented training environments for the data-intensive algorithms since early phases of the development process, simulation methodologies using X-in-the-Loop (XiL) platforms lend themselves well. The work at the institute spaces from the conceptual design of the function to the automation of the training process though different levels of the XiL platform, to the implementation on the real vehicle.