| Resumo: | A soccer-playing robot must be able to carry out a set of behaviors, whose complexity can vary greatly. Manually programming a robot to accomplish those behaviors may be a difficult and time-consuming process. Automated learning techniques become interesting in this setting, because they allow the learning of behaviors based only on a very high-level description of the task to be completed, leaving the details to be figured out by the learning agent. Reinforcement Learning takes inspiration from nature and animal learning to model agents that interact with an environment, choosing actions that are more likely to lead them to accumulate rewards and avoid punishment. As agents experience the environment and the effect of their actions, they gain experience which is used to derive a policy. Agents can do this instantaneously after they observe the effect of their last action, or after collecting batches of these observations. The latter alternative, called Batch Reinforcement Learning, has been used in real world applications with very promissing results. This thesis explores the use of Batch Reinforcement Learning for learning robotic soccer behaviors, including dribbling the ball and receiving a pass. Practical experiments were undertaken with the CAMBADA simulator, as well as with the CAMBADA robots.
|
|---|