Using ensemble of decision trees with SVM nodes to learn the behaviour of a transmission control software

Energy efficiency has become a major issue in trade, transportation and environment protection. While the next generation of zero emission propulsion systems are still under development, it is already possible to increase fuel efficiency in regular vehicles by applying a more fuel efficient driving behaviour. The ensemble classifier presented in this paper is part of an adaptive manufacturer independent fuel efficiency assistant that only uses publicly available FMS-2 CAN-Bus data. The goal is to learn the basic behaviour of an unknown automatic transmission control software, only by investigating available input and output data. The knowledge can then be used to e.g. predict the fuel consumption of a vehicle or be used for other purposes that are not subject to this specific paper. The classifier consists of random ensembles of global and local classification trees, whose nodes are binary Support Vector Machines. To the best knowledge of the authors, it is the first time this specific kind of classifier has been formulated and used to learn the behaviour of an unknown software based on rudimentary input and output data.