Generation of motion and decision-making policies applying multiagent reinforcement learning in simulated robotic soccer.

Abstract

When designing intelligent agent systems, it is impossible to implement all the potential situations an agent may encounter and specify an optimally behavior in advance. An intelligent agent must be capable of using its past experience to learn how its actions affect the environment in which is situated. Reinforcement learning (RL) is a machine learning technique where an agent learns from its own environment while the agent is in it. With this technique, goal directed agents learn what to map states to actions in order to maximize their utility. R L techniques have addressed the problem of learning for a single agent acting in a stationary environment and for many agents acting in a dynamic environment. Multiagent environments are inherently dynamic since the agents may change the environment simultaneously, as all of them are learning and adapting at the same time. This thesis aims to implement both single agent and multiagent R L approaches in the international "RobotStadium" simulated robotic soccer competition in order to generate motion and decision-making policies. The "RobotStadium" competition is a simulated soccer league based on the "Nao" robot (from Aldebaran robotics) and the "Webots" simulator (from Cyberbotics), using the same rules of the Standard Platform League (SPL) of the international "RoboCup" competition. The "RobotStadium" platform offers an excellent experimental framework for multiagent systems under a realistic, uncertain and highly dynamic environment.