This repository is home of a Machine Learning project aiming at training an agent to play a game using Reinforcement learning
2 personnes pour :
Créer un jeu simple (casse-brick) qui communique avec un script python. (Kevin and Matthieu)
2 personnes pour :
Etudier un cas existant d'apprentissage par renforcement (Nicolas and Benjamin)
POC
And ... we didn't push it further! I mean the planning !
Then it was about updating the model to fit our game.
So, for the breakout-game, we had two different representations :
The first one was an array of three boolean which help giving the ball position. The results with this were quite mediocre and we agreed on a change.
So we decided to do as in the article we were starting from. We send the whole screen shot of the game to our network. Maybe not exactly the screenshot, but a reduced version of it. By doing so, we may have a lot of the informations a real player has.
That's what you can find in the game folder.
We adapted a game to our convienence and used it to train on it. the q_learning script mais train and test 2 differents games: our breakout adn the gym[atari] breakout
To launch the script, here is a little guide:
python q_learning.py [game] [action] (--model path_to_the_model_to_restore)
game : Required. The game you want to launch. It have twos valid values for the moment : atari, homemade
action : Required. What we want to do with the game. Also twos valid values : train and test
path_to_the_model_to_restore : Required for testing, optional for training. Path to the model we want to restore in our network
Well, the snake game was the first example we found on the net of a working q-learning algorithm.
It was kind of a side quest for us. Turns out to be more difficult than accounted for but here we are.
You will find more detailled informations on how to use it in this document