Fuzzy Controller Inference via Gradient Descent to Model the Longitudinal Behavior on Real Drivers

This paper introduces a method to represent Takagi-Sugeno Fuzzy Control Systems (FCSs) as computational graphs, so they can be adjusted through a supervised training process based on gradient descent. It has been tested both with artificial (i.e. a known fuzzy controller) and naturalistic (i.e. driver's data extracted from the vehicle and the environment) data. The results achieved show high conformance to synthetic data, and seem to describe a car-following behavior with quite good precision, which suggests that it is possible to model the driver's behavior in a longitudinal model based on if-then type rules.