RL-Project TP4

Deep Reinforcement Learning project on the Pong game developed by:

• Juliette Jacquot
• Matis Braun

Report

The report is the pdf Rapport_RL_Pong_Project.pdf on Github.

Deep Q Learning:

If you want to use the Deep Q Learning model, you have to launch the main.py, which will launch the training on 1,000,000 frames. This model has the worst results.

Double Q Learning:

If you want to use the Double Q Learning model, you have to launch the main_ddqn.py, which will launch the training on 2,000,000 frames. This model has the best results. The best model has been saved in the file model_ddqn_best.joblib

Dueling Double Q Learning + Noisy Network

If you want to use the Dueling Double Q Learning + Noisy Network model, you have to launch the upgrade/main_ddqn.py which will launch the training on 3,000,000 frames. This model theoretically has the best results.

Animation

In the animation folder, we can find best_final_game_ddqn.gif which shows a game using the best model we had. The model wins with a 19-point lead.

In the animation folder, we can find strange_win.gif which shows the game where the model managed to win without moving the whole game, which is impressive.