Ilham Saputra, Adhan Efendi, Efrem Olivio Gomes, Yohanes Sinung Nugroho
This study investigates the application of Deep Q-Learning (DQN) to optimize the performance of agents in the Flappy Bird game environment. By leveraging the principles of reinforcement learning (RL), the agent learns to make decisions based on state-action values, updated iteratively throughout the training process. The research methodology involves hyperparameter testing to evaluate the model's effectiveness. Training results reveal a consistent decrease in the loss function, accompanied by gradual improvements in Q-values, signifying enhanced decision-making capabilities of the agent. The reward signal initially stabilizes at 0.2, with occasional spikes reaching 1.1, reflecting the agent's ability to identify and execute more optimal actions. Notably, the trained agent successfully navigated over 80 consecutive pipes without failure, demonstrating the model's high stability and learning efficiency. This study underscores the potential of DQN in addressing dynamic, continuous-state environments and sets a foundation for applying reinforcement learning to similar arcade-style games. Future work could focus on enhancing the model's performance through advanced techniques, such as Double DQN and Prioritized Experience Replay, to reduce overestimation bias and improve sampling efficiency. © 2025 IEEE.
National Yunlin University of Science and Technology, Department of Mechanical Engineering, Douliu, 64002, Taiwan; Universitas Negeri, Faculty of Engineering, Department of Mechanical Engineering Education, Surabaya, Indonesia; Politeknik Negeri Bandung, Department of Aeronautical Engineering, Bandung, 40559, Indonesia
