Playing Atari Games using Deep Reinforcement Learning

This is the final year project for Team Convolution. We have done a survey about Deep Reinforcement Learning and are testing its applications on various Atari Games.

We are trying to build a Reinforcement Learning Agent for atari game using Asynchronous Advantage Actor-Critic (A3C) algorithm which has been described in this paper.

This code is heavily inspired from the works of OpenAI/universe-starter-agent and Deep-RL agent. You may go through these codes if you feel like doing so.

We have implemented the A3C algorithm and have tested the various Gradient Descent/Ascent Optimization Techniques like Adadelta, RMSProp and Adam. You can read about them here.

Running the code

Firstly make sure you run script_install_before.sh in your terminal so that all the prerequisite libraries are installed.

To run the code, please check the file run.sh. Each command in that file needs to run on a separate terminal. A terminal manager like tmux can also be used.

To see the progress, open http://localhost:6006/ in your browser.

Team Members

Piyush Bhopalka, Saksham Agarwal, Mahesh Uligade

Please shoot an email at [email protected], if you have any queries :)

(National Institute of Technology, Calicut)