ABSTRACT
This paper presents a multi-agent control architecture for the efficient control of a multi-wheeled mobile platform. Such control architecture is based on the decomposition of a platform into a holonic, homogenous, multi-agent system. The multi-agent system incorporates multiple Q-learning agents, which permits them to effectively control every wheel relative to other wheels. The learning process was divided into two steps: module positioning – where the agents learn to minimize the error of orientation, and cooperative movement – where the agents learn to adjust the desired velocity in order to conform to the desired position in formation. From this decomposition, every module agent will have two control policies for forward and angular velocity, respectively. Experiments were carried out with a simulation model and the real robot. Our results indicate a successful application of the purposed control architecture both in the simulation and in real robot.
