Optimization of system parameters for an online driving style recognition

An online driving style recognition system using fuzzy logic has recently proven to work well and showed potential to optimize its parameters. This paper is about the efficient parameter optimization of such a system. To overcome combinatorial explosion, we introduce heuristics to express the main influential parameters of the system, which itself is divided into two layers. First, we use a method called Design of Experiments in order to identify the most important parameters of general high-level system parameters. The low-level layer consists of fuzzy logic systems, which are the core of the driving style recognition system. For this, we introduce a way to efficiently describe the main characteristics of a fuzzy system by very few parameters. Both sets of identified parameters are then separately optimized with an established multidimensional evolutionary algorithm. We show that using Design of Experiments is superior to a random selection of the high-level parameters, as it increases the optimization gain by 76.5% in average. All in all, the target function, which represents a weighted classification error, was reduced by 43.9% on the test data set. The optimization method can be used to calibrate the system on real-world driving data. The combination of Design of Experiments, evolutionary optimization and fuzzy logic parametrization can also be used to optimize arbitrary other complex nonlinear systems.