sinarazi / FlapAI-Bird

An AI program that plays Flappy Bird using reinforcement learning.

FlapAI-Bird

This AI program implements several AI agents for playing Flappy Bird. The program applies reinforcement learning algorithms, including SARSA, Q-Learning, and Function Approximation, and Deep Q Networks. After training for 10,000 iterations, the agents regularly achieves high scores of 1400+, with the highest in-game score of 2069. [Paper] [Poster]

Installation

The project requires the following frameworks:

• Pygame: https://www.pygame.org

• PyGame Learning Environment: https://github.com/ntasfi/PyGame-Learning-Environment

• OpenAI Gym: https://gym.openai.com

• gym-ple: https://github.com/lusob/gym-ple

• PyTorch: https://pytorch.org

• OpenCV: https://opencv.org

Train an Agent

The program implements the following agents.

• Baseline Agent.

python main.py --algo Baseline --probFlap 0.5

• SARSA Agent.

python main.py --algo SARSA --probFlap 0.1 --rounding 10 --lr 0.8

• Q-Learning Agent.

python main.py --algo QLearning --probFlap 0.1 --rounding 10 --lr 0.8 --order backward

• Function Approximation Agent with Linear Regression.

python main.py --algo FuncApproxLR --probFlap 0.1 --lr 0.1

• Function Approximation Agent with a Feed Forward Neural Network.

python main.py --algo FuncApproxDNN --probFlap 0.1 --lr 0.1

• Function Approximation Agent with Convolutional Neural Network.

python main.py --algo FuncApproxCNN --probFlap 0.1 --lr 0.1

Authors

• Tai Vu - Stanford University

• Leon Tran - Stanford University