Using the GTSOM network for mobile robot navigation with reinforcement learning

This paper describes a model for an autonomous robotic agent that is capable of mapping its environment, creating a state representation and learning how to execute simple tasks using this representation. The multi-level architecture developed is composed of 3 parts. The execution level is responsible for interaction with the environment. The clustering level, which maps the input received from sensor space into a compact representation, was implemented using a growing self-organizing neural network combined with a grid map. Finally, the planning level uses the Q-learning algorithm to learn the action policy needed to achieve the goal. The model was implemented in software and tested in an experiment that consists in finding the path in a maze. Results show that it can divide the state space in a meaningful and efficient way and learn how to execute the given task.