| Summary: | Throughout the years the RoboCup, an international competition of robotics and artificial intelligence, saw many developments and improvements in these scientific fields. This competition has different types of challenges including a 3D Simulation League that has an annual tournament of simulated soccer games played between several teams each composed of 11 simulated humanoid robots. The simulation obeys the laws of physics in order to approximate the games as much as possible to real circumstances, in addition, the rules are similar to the original soccer rules with a few alterations and adaptations. The Portuguese team, FC Portugal 3D has been an assiduous participant in this league tournaments and was even victorious several times in the past years, nonetheless, to participate in this competition is necessary for teams to have their agents able to execute low-level skills such as walk, kick and get up. The good record of the FC Portugal 3D team comes from the fact that the methods used to train the robots keep being improved, resulting in better skills. As a manner of fact, it is considered that these low-level behaviors are at a point that is possible to shift the implementations' focus to high-level skills based on these fundamental low-level skills. Soccer can be seen as a cooperative game where players from the same team have to work together to beat their opponents, consequently, this game is considered to be a good environment to develop, test, and apply cooperative multi-agent implementations. With this in mind, the objective of this dissertation is to construct a multi-agent setplay based on FC Portugal's low-level skills to be used in certain game situations where the main intent is to score a goal. Recently, many 3D League participants (including the Portuguese team) have been developing skills using Deep Learning methods and obtaining successful results in a reasonable time. The approach taken on this project was to use the Reinforcement Learning algorithm PPO to train all the environments that were created to develop the intended setplay, the results of the training are present in the second-to-last chapter of this document followed by suggestions for future implementations.
|
|---|