Playing Atari with Deep Reinforcement Learning: Comparing Individual Models with Ensembles and Soups

This repository accompanies my master's project at Ulm University. It contains the code for training and evaluating the models, as well as plotting the results of the experiments.

I used Gymnasium for the RL environments and trained the models following the 2013 DQN paper and 2015 follow-up paper. Afterward, I created ensembles and soups of the trained models and compared them to the individual models. You can take a look into the written report here.

• Training the models: src/Training
• Evaluating the models: src/Evaluation
• Plotting results and doing significance tests: src/Visualization

Everything related to latex and all plots are in the latex/ folder. The Notebooks that were used to generate the plots automatically save them to the latex folder.

All trained models and the results of the evaluation (>9GB) can be downloaded here: Cloudstore Uni Ulm