Automatic car parking: A reinforcement learning approach

Abstract

The automatic parking of a car-like robot is the problem considered in this paper to evaluate the role played by formal representations and models in neural-based controllers. First, a model-free control scheme is introduced. The respective control actions are sensory-based and consist of a dynamic, neural-based process in which the neurocontroller optimizes ad hoc performance functions. Afterwards, a model-based neurocontroller that builds without supervised a formal representation of its interaction with the environment is proposed. The resulting model is eventually utilized to generate the control actions. Simulated experimentation has shown that there is an improvement in robot behavior when a model is used, at the cost of higher complexity and computational load. © Springer-Verlag Berlin Heidelberg 2003.