Öztürk A., Noyan Tekeli F.

8th International Conference on Advances in Statistics, Gazimagusa, Cyprus (Kktc), 16 May - 18 June 2022, pp.29

  • Publication Type: Conference Paper / Summary Text
  • City: Gazimagusa
  • Country: Cyprus (Kktc)
  • Page Numbers: pp.29




Reinforcement learning (RL) is currently one of the most dynamic areas in Artificial Intelligence research. It is the technique in which an agent learns how to obtain rewards through interactions with their environment. Some real-world applications, such as robotics and autonomous, are particularly suitable for the RL approach as the environment is unknown and the consequences of actions are uncertain. Because many Reinforced Learning examples in the literature are trained with simple and intuitive reward functions, it struggles with sparse and deceptive feedback. In addition, it may be necessary to search for different ways for trial and error and sometimes return to the previous steps after discovering these ways, so it is important to use the right search algorithms. Discovering all paths in remains one of the challenges of the field. In Reinforced Learning, it may not be the optimum solution to give high values to the activities we want to take place or to apply a penalty system to undesirable activities. However, there is no proven method that selects the appropriate reward function for every situation. The reward functions used are intuitive. In this study, it is planned to train the Reinforced Learning algorithm in a game environment by giving different reward functions. It is aimed to examine the performances and strategies of the learning algorithms and to make comparisons between them. However, for this, it is necessary to have environments where we can directly access the situations and actions in the game, and the diversity of decisions that lead to the result in the game for the reward function. DOOM 2 game is suitable for this study is found by examining the environments in the literature. In this game there are multiple chapters and different obstacles that will ensure that the chapter designs do not reach the goal directly. It is aimed to compare the effects of department diversity and barriers on education, as well as to examine the performance and strategies of representatives trained in the same departments with different reward functions and to make comparisons between them.


Key Words: deep learning; reinforcement learning; video games; deep q learning