FlappyAlpha: Reinforcement Learning vs Human

🏆 This is my project inspired by AlphaGo's groundbreaking success in reinforcement learning (RL). I developed a custom Flappy Bird game environment featuring realistic physics like gravity, where an RL agent (I named it as FlappyAlpha) learns to navigate by itself based on a reward and punishment system.

🏆 The agent receives positive rewards for passing obstacles and negative penalties for hitting the top, bottom, or crashing into pipes, which results in "death." I trained the RL agent across multiple iterations—2500, 3000, and 4000 episodes—to analyze how its performance evolves with increasing learning experience.

🏆 As the number of training episodes increased, the agent became progressively adept at "mastering" the game based on experience. After 3000 episodes compared to 2500, as the number of playing times increases by 20%, the agent became more familiar with game environment and improved its game scoring rate, with the best score soaring from 17 to 75—a remarkable 441% increase. By 4000 episodes, the agent's best score reached an astonishing 3693, demonstrating near-perfect dominance over the game.

🏆 To showcase its capabilities, I pitted human players against two versions of the agent: a 2500-episode "Beginner" version and a 3000-episode "Hard" version. In the Beginner version, humans easily outperformed the agent. However, the Hard version, after 3000 episodes, proved much tougher—over 5 rounds, humans narrowly won 3 games, but the agent’s average score surpassed human performance, which is incredible. I didn’t release a 4000-episode version for human challenges, as internal unofficial tests showed the agent effortlessly achieving scores of 100+ with games lasting significantly longer, making it impractical for human competition—but the results speak for themselves: the RL agent has achieved extraordinary mastery.

Tech Stack
- Game physics simulation and design: Python, Pygame
- Reinforcement Learning algorithms: Q-Learning
- Data Analysis: NumPy, Matplotlib