Heavy vehicle path control with neural networks - Heavy vehicle path control with neural networks

Abstract

This thesis explores the possibility of using neural networks for solving the path control problem, i.e. how to follow a predefined path as closely as possible. Two main approaches are used to achieve this, namely supervised learning and reinforcement learning. The supervised learning approach is based on existing path trackers which are used to generate data for the training procedure. The reinforcement learning uses a genetic algorithm and simulations to evaluate possible solutions. The supervised learning controllers are constructed as feed forward neural networks only, while the reinforcement learning controllers uses a recurrent neural network. The results shows that neural networks can be trained to solve the path tracking problem, both with supervised and reinforcement learning methods. Both the feed forward networks and the recurrent networks outperform the geometric path trackers. Further, a recurrent network was shown to perform better than a feed forward network, which indicates that the dynamical properties of such networks can be useful in path tracking applications.