This set of experiments originates as a project work for the Deep Learning module, held at University of Bologna. See the complete report: Matering the game of Corso without human knowledge.
This directory contains all the artifacts obtained from the experiments, and the specific scripts used to do so. In particular:
5x5_64x4/contain the artifacts relative to their algorithm. That is:- The pretrained model
az_network_5x5_64x4. - All the related
tensorboardlogs (runs/directory). - A parquet dataset (
runs.parquet) containing a tabularized summary of all the runs. - Figures generated for the final report (
*.svg).
- The pretrained model
figure_scripts/contains the scripts used to generate all the figures. These are rough and undocumented due to time constraints, in the end it is nothing more than some data manipulation in order to obtain clean figures.figure_requirements.txtcontains the necessary requirements to run them (pip install -r figure_requirements.txt).tournaments/contains numpy arrays representing the results of tournaments played between a pool of agents and with different seeds.other_figures/, as suggested by the name, contains other figures not specifically related to an algorithm/agent.seeds.txtcontains the only seed used during training. NOTE: reproducibility is not yet properly implemented.train5x5_64x4.pyis the driver script which runs the training pipeline with the manually tuned hyperparameters. The only requirement to run it is a correct installation of thecorsolibrary.