Constrained Training for Guided Feedback Controller

Abstract

In this paper, we are focusing on a neural network feedback controller tailored to a lane following application. Besides a classical goal like avoiding a collision with obstacles, we aim at investigating an approach in order to guarantee not to leave the street at any point, since this is of major importance for the safety of the passengers. Unfortunately, guarantees like this are quite rare since the training of neural networks is often based on data, which only allows statistical conclusions. We derive a deterministic inequality equation guaranteeing this safety. Furthermore, we introduce a penalty strategy in order to integrate this inequality into a Reinforcement Learning algorithm.