Using a genetic algorithm to obtain a neural network-based model of a real autonomous vehicle

In this paper, a set of radial basis function (RBF) neural networks, capable to learn the kinematic and dynamic behavior of the Romeo 4R autonomous vehicle, is presented. In order to obtain a set of good RBF nets in terms of the number of neurons and the number of lagged inputs, a multi-objective genetic algorithm (MOGA) has been used. The kinematic and dynamic systems of the mobile robot have been split into three subsystems: the steering module, the drive module and the heading module. Each subsystem is modeled with a neural network that learns its behaviour using, among others, a set of lagged outputs as inputs. The outputs from the steering and drive modules are also used as inputs in the heading module. Neural networks - based models are compared to classical approaches.