atari_rl

This repository includes trained dqn agent to play Breakout (Atari Game) using Keras/Tensorflow and Open Gym AI.

This repository includes:

• train_agent.py - python script to training agent. Training can be paused and resume. Score during training and other parameters are saved to log_training directory. Trained models are saved to trained_network directory.
• test_agent.py - python script to test trained agent. Testing can be recorded and saved as *.mp4 file. Results of testing are saved to log_test.txt file.
• training_graph.py - python script to generate graph from training log files.

Requirements:

• Python 3.7.4,
• Keras, Tensorflow,
• Open Gym AI (Atari version),
• Numpy, Matplotlib.

Agent description

Agent was learned by 37 hours (about 9k games). Score during training is shown on the bottom:

Agent progress during testing (training after 5h, 7h, 18h, 22h and 37h):

Comparison between maximum training score and maximum testing score in different training time is shown below: