Hendrickx, JulienStandaert, François-XavierVermeulen, CorentinCorentinVermeulen2025-05-142025-05-142025-05-142024https://hdl.handle.net/2078.2/37632Artificial Intelligence (AI), specifically reinforcement learning (RL), has significantly advanced through the development of Deep Q Networks (DQNs). DQNs effectively combine Q-learning with deep neural networks to manage high-dimensional state spaces. Whilemost studies focus on DQN algorithm architectures, this thesis inves- tigates the impact of different environmental configurations on learning complexity and performance. The research involved training a DQN agent in various environmental settings to assess its learning efficiency and performance. The environments included both random and fixed conditions, varied action spaces, and different levels of uncertainty in action execution. The Flappy Bird game environment, implemented via OpenAI Gym, served as the testbed for these experiments. The findings indicate that DQN agents trained in random conditions outperformed those trained in fixed conditions, demonstrating better generalization capabilities. Enlarging the action space was observed to slow down learning and reduce per- formance. Furthermore, introducing uncertainty in action execution had minimal impact on learning velocity when the uncertainty did not alter game dynamics. However, significant changes in game dynamics due to uncertainty drastically reduced learning speed and performance. The study concludes that environmental configurations substantially influence the learning and performance of DQN agents. Training in varied conditions en- hances generalization, while increased action spaces and dynamic uncertainties pose challenges to learning efficiency. These insights can guide the development of more robust reinforcement learning systems capable of adapting to diverse real-world environments.Deep Q LearningReinforcement learningVideo gamestext::thesis::master thesisthesis:45972