Título: Experiments in Robot Control for an Instance-Based Reinforcement Learning Algorithm based on Prior Information
Autores: Ribeiro, Carlos H. C.; Hemerly, Elder M.
Resumo: Reinforcement learning techniques are based on experience acquisition. In most realistic applications, experience is time-consuming: it implies sensor reading, actuator control and algorithmic update, constrained by the learning system dynamics. In fact, the information crudeness upon which classical learning algorithms operate can make such problems too difficult and unrealistic: embedding of prior structural knowledge, learning of control laws (instead of low level control actions) and full use of experience are required if control of real systems is at stake. In this paper, we show how a robust formulation (with respect to convergence properties) of the Q-learning method that considers prior information about the structure of the state space can be combined with an instance-based technique as a mechanism to accelerate learning. We then demonstrate it both in simulation (with a statistical consistency test) and in a real robot for a guidance task defined by a combination of a predefined control law and learned action policies.
Artigo em pdf: 4cbrn_009.pdf
Arquivo BibTex: 4cbrn_009.bib