- Lieuwe Meijdam (s1047496)
- Marius M. Kästingschäfer (s3058301)
- Djesse Dirckx (s1046968)
This repository contains all software that was used for the final project of the Natural Computing course. In this project, it was tried to implement the World Model architecture for the game MsPacMan. Specifically, it was evaluated which approach (genetic algorithm vs. evolutionary strategy) works best for optimising the parameters of the controller component. The following sections describe requirements and files that are needed to replicate results.
All code in this repository is written in the Python programming language. To be able to succesfully run the experiments, the following dependencies are required:
atari-py==0.2.9
cma==3.0.3
gym==0.18.3
h5py==2.10.0
notebook==6.1.4
matplotlib==3.3.2
numpy==1.19.2
opencv-python==4.5.1.48
pandas==1.1.3
scikit-image=0.17.2
scikit-learn==0.23.2
tensorflow==2.4.1
For running the Atari environment, a ROM file containing the game dynamics is required. This can be found here.
The steps below can be followed for replicating the VAE and RNN experiments:
- The data required for training the VAE can be generated using the Data collection for VAE notebook.
- To use the RNN, either the listed pretrained models or the notebook for training the RNN can be used for retraining the model.
In order to repeat the various experiments to find an optimal strategy for the agent, the respective notebooks listed under Files
can be used. It is important to execute these notebooks from top to bottom in order to get proper results. All experiments can be conducted independent from each other. Each notebooks contains detailed comments about its components.
The files in this repository are distributed over multiple folders
- Images:
- RNN weights
- Source code:
- SQM
- GA
- CMA-ES