Bachelor Thesis - 'Hide and Seek Games'

Reinforcement learning is considered a dominant approach to decision-making in dynamic and uncertain environments. One field of usage for reinforcement learning is in pursuit-evasion games, where traditionally a pursuer tries to follow and ultimately catch an evader that is trying to avoid the pursuer. Countless different pursuit-evasion games have been studied extensively in the context of game theory, robotics, and artificial intelligence. Reinforcement learning algorithms have been shown to be effective in training agents in pursuit-evasion games. These algorithms can be applied to a wide range of problems, including autonomous drone navigation, military operations, and search and rescue missions. In this thesis, the method of training agents using curriculum learning which was introduced by Qi et al. will be reconstructed and tested thoroughly. Furthermore, the pursuit-evasion game will be extended to an environment entailing multiple pursuers simultaneously, eventually using the MADDPG algorithm by Lowe et al. to increase the training efficiency.