The main concepts of machine learning are presented in the thesis with an emphasis on reinforcement learning. The latter area also covers problems with multi-agent environments. Such problems pose additional challenges to the methods of reinforcement learning. The thesis explores various ways to solve multi-agent problems, using prediction of adversary actions in intelligent agent learning. In the diploma work, the DRON method, which is designed on the basis of deep q-learning, is presented in more detail. It is compared with the basic method of deep q-learning on the selected environment. The work also presents and compares the extension of curiosity-driven exploration methods to a multi-agent environment. The environment was implemented in the game engine Unreal Engine 5. In the end, the presented exploration method did not prove to be noticeably more successful than the basic exploration method. On the other hand, the use of the DRON architecture in combination with the basic deep q-learning method indicated a potential improvement of the basic method. For more concrete conclusions, additional experiments should be carried out.
|
