Adaptive situation-based learning of a dribbling behavior in RoboCup

In order to test and rate any algorithm, test problems, called benchmarks, are needed. Sorting algorithms are evaluated using random or in some way presorted datasets to understand their performance in different scenarios. The same has to be done with algorithms in artificial intelligence. For algorithms which try to simulate human thinking in general, the most famous benchmark is the Turing test, in which the system has to fool a user who is asking questions to think the responding unit is in fact human. To date (2011) no system has come close to passing the test consistently. There are however other domains of AI research which can not be as easily tested as the capability of a system or algorithm to respond to questions like a human. These include cooperation, learning and sensor and actor technology. The RoboCup is a yearly competiton where exactly these domains are benchmarked in soccer games between teams from universities from all around the world. The rules of the contest are very close to the official soccer rules stated by the international soccer federation FIFA. Some adaptations are necessary to make participance in the competition affordable and to rid the game of elements which are unnecessary when robots are playing, like a dedicated referee who is present on the field. There are five different leagues with different rules. The university of Stuttgart takes part in the Middle-Size league with non-humanoid robots of a maximum of 50cm in diameter and a height of approx. 1m. They play in teams of five robots each with standard soccer balls on a field which is scaled to a length 18m, but otherwise nearly identical to the standard soccer field outlined by FIFA regulations. The main focus of this competition is on cooperation and planning. Other leagues have other regulations and a different focus. What makes the RoboCup scenario so valuable for research in Artificial Intelligence is its focus on a dynamic environment and incomplete information accessibility, that means the input of the sensors of the robots is never entirely certain. The principles and methods used to compete in the RoboCup are useful in other domains and scenarios. In this thesis an algorithm to adapt behaviour based on the uncertain sensor input is developed and introduced, as well as tested. The input is provided by an omnidirectional camera on top of the robot and by referee events delivered to all of the robots during the game. The output of the algorithm, that means the behaviour of the robot is (in certain bounds) accessible by parameters, which for example control the shooting mechanism. So the whole system can be represented as a control system where the robot is the control process and our algorithm is the controller. Common controller design methods fail here, the special difficulties of this system lie in the incertainty of the input signal and the highly non-deterministic environment. For the design of a controller for this kind of process, new methods and even controller components have to be developed