Machine Learning Techniques Applied to Multi−Agent Cooperation
Julian de Hoog
Recent rapid advances in technology mean that robots are being used for an ever widening range of tasks. While traditional robotics tasks have typically involved robots operating individually, there is now an increasing need for development of techniques that may be useful for cooperation between robots, in other words for multi-agent cooperation. This field requires robust and versatile solutions since robots must typically operate in noisy and uncertain environments. Robot football is an increasingly popular academic discipline worldwide. The challenges offered by robot football are numerous, and solutions to robot football problems often apply to a wider range of robotics applications. This makes robot football an ideal testbed for the study of multi-agent cooperation problems. In this thesis, two areas of multi-agent cooperation are approached within the framework of robot football: collision planning and path planning. Artificial neural networks are applied to specific problems in each of these areas, and in both cases lead to an increase in performance. In the case of path planning, a novel approach is proposed. All experiments are undertaken within the simulator developed for the FIRA robot football simulation league. The results of these experiments indicate that machine learning techniques are well suited to solving multi-agent cooperation problems. Certain machine learning techniques are particularly suitable for certain types of problems. Machine learning techniques outperform non-learning techniques in a number of cases, and there is extensive scope for machine learning techniques to be applied to a wider range of collision and path planning problems, and to multi-agent collaboration problems in general.