Building my own DQN-Atari-Agent Library

In this upcoming series, I want to write about my journey of building my own modularized library to train DQN to play Atari games. The plan is to make it very easy to test and train several different Agents. With one base code script, the user shall be able to switch easily between the different addons for DQN. Simply by changing the arguments of the main script.

Starting with the base algorithm DQN the addons shall go on with for example: Dueling DQN, C51, PER to Rainbow and end with DRQN.

My goal is to make it easy to track experiments with all those algorithms and to see how the algorithms evolved and improved over time.

This implementation and documentation will be ongoing and you can check the progress directly on Github or here on Medium.

The Agents will be implemented with Pytorch and the Atari environments are provided by the gym library of OpenAI. My plan is to be as close to the original descriptions in the papers and their implementations.

___________________________________________________________________

Disclaimer: This and the following articles are no tutorials and just a journal that describes the work and progress that has been done. A certain level of understanding of Deep Learning and Deep Reinforcement Learning is needed, if you…