bingo-hyouka / dqn-atari-pong

This project explores a deep reinforcement learning technique to train an agent to play atari pong game from OpenAI Gym. OpenAI Gym is a toolkit to develop and compare reinforcement learning algorithms. The learning agent takes raw pixels from the atari emulator and predicts an action that is fed back into the emulator via OpenAI interface. The deep reinforcement learning network used in this project is Deep Q Network (DQN), it took over 10 million episodes to perfectly play and win the game.

Machine Learning Nano Degree Capstone Project: Deep Reinforcement Learning to train an agent to play Atari Pong game using DQN

Random model	DQN 5260000 steps	DQN 8090000 steps	DQN 9500000 steps

I have obtained the starter code from CS 294: Deep Reinforcement Learning, Spring 2017

https://github.com/berkeleydeeprlcourse/homework/tree/master/sp17_hw/hw3

Make sure to install Tensorflow, OpenAI Gym, numpy, pandas and bokeh before running these scripts.

Run the training using the following command

python run_dqn_atari.py >> train_logs_new.log

There is a jupyter notebook, named DQN report.ipynb, which has all the code for plotting the stats presented in the report. The file named DQN_report.html is the html snapshot of the jupyter notebook. The file named record.pdf has the complete report of the project.