This project is a modified version of the original FlapPyBird game, which now includes basic AI agents designed to autonomously play the game. Modifications to the original game include the addition of a get_state() method alongside cool slow-motion and fast-forward features to enhance the fun of training and testing.
The AI started with a QDN (Q-Network Deep Neural) agent but later transitioned to a combination of a Reflex-approved agent and a Q-Learning agent. This hybrid approach enables the system to identify Killer Moves when possible, and otherwise, it relies on the Q-Learning agent to derive the optimal game-playing policy.
The Reflex agent employs a simple heuristic, primarily based on the positions of the bird and the pipes, as well as the bird’s velocity. It searches for potential Killer Moves to navigate the game's challenges effectively.
Our Q-Learning agent experiments with various reward functions to optimize performance. After rigorous testing, including time-based rewards and distance-based rewards, the most effective metric has been the absolute distance from the bird to the closest pipe's middle point.
If you have any ideas or need further information, please feel free to contact me via my LinkedIn profile: Dor Pascal.
Credit for the original FlapPyBird game goes to its creator, sourabhv. Please refer to the original game at the link provided above. This project is a derivative and does not claim original ownership of the underlying game concept.