In this project, a reinforcement learning agent was trained to solve a classic video game 'Space Invaders'. The reinforcement learning used in this project is Deep Q Network.
In order to play the variant of a classic video game Space Invaders, you will need to install the dependencies first. Run the following command in your terminal:
pip install -r requirements.txt
pygame, keras, tensorflow, pickle, pyautogui
This project involves mainly three components: the video game (several py files), the learning agent (dqn_agent.py) and the API which operates as the interface between the game and the agent. The video game scripts are built with several .py files which are implemented with the help of pygame, the API was implemented with pyautogui, which allows the agent to 'actually' play the game, and last but not least, the learning agent was implemented by using the high-level deep learning framework keras, and the learning algorithm used is Deep Q Network.
The reward function is designed as the score gained during the game. When the bullet hits an alien, we or the agent get 50 points. The agent has to learn how to react (go left or right) based on the state at that time. In this game, there are three discrete actions available.
left arrow- move leftright arrow- move rightspace key- fire a bullet (limit: 3 bullets at the same time)
After the dependencies are installed, you can first try out the game by running the
file space_invaders.py.
python space_invaders.py
You can see how the agent is being trained by running the file main.py.
python main.py
The agent is able to achieve the same high score as we human can. We limited the playing time to 100 game cycles, which is approximately equal to 14 seconds. The final highest score the agent achieve is 450, which is the same as what I can do with 10 trials.