Q-learning and robotics

Abstract

Because it allows the synthesis of behaviors despite theabsence of a robot-world interaction model, Q-learning hasbecome the most used learning algorithm for autonomousrobotics in applications such as obstacle avoidance, wallfollowing, go-to-the-nest, etc. This is mostly due toneural-based implementations such as multilayer perceptronstrained with backpropagation, or self-organizing maps. Suchimplementations provide an efficient generalization, i.e.,fast learning, and designate the critic, the reinforcementfunction definition, as the real issue. The paper discussesQ-learning for robots and Q-Kohon self organising map.