An implementation of the Deep-Q algorithm for Reinforcement Learning problems as described in the papers Playing Atari with Deep Reinforcement Learning and Human-level control through deep reinforcement learning.
The algorithm has proven to be able to generate a policy that solves most Atari games. In this implementation we focus on the game Pong.
A trained policy that achieves an approximate average return of 5.8 is provided in the models folder. You can run it using
python evaluate.py --env Pong-v0 --path models/Pong-v0_best.pt
Note that there is some volatility in the performance over several runs.To get an estimate of the expected mean return you can additionally set e.g. n_eval_episodes=10. You can also specify render if you want to see the agent in action.
Model training can be performed by running
python train.py --env Pong-v0
However, it should be noted that the model training is time-consuming. If you don't have a GPU at hand, I recommend performing the model training using Google Colab and make use of their GPUs. A notebook with instructions on how to do so is provided in DeepQLearning.ipynb.
The code is the result of a project as part of the course Reinforcement Learning that I took at Uppsala University in the spring 2021.