This is the full repository for the project titled A Deep Q-Learning approach to Dino Run created for the Decision Models course @ Università degli Studi di Milano Bicocca A.Y 2019-20.
Refer to the report.pdf
file for an in-depth methodological explanation.
Due to its high complexity, it is impossible to train the model on regular hardware. Even using a pre-trained model, the latency is too high on commodity hardware to be able to communicate efficiently with the web browser.
For this reason, a Google Colaboratory notebook was prepared, ready to run on a cloud Linux VM. To access it, click here. The notebook is well documented and nothing else is needed to run the code.
The model/model.h5
file is an HDF5 snapshot of the most recent model, and will be used in the notebook to ensure replicability.
The Colab notebook is self-contained. If you instead insist on running locally, be careful as the program depends on a Chromedriver executable to interact with the web browser.