This project implements the experiments for the paper Illuminating Mario Scenes in the Latent Space of a Generative Adversarial Network. Our project implements a method for Latent Space Illumination (LSI) which explores the latent space of generative adversarial network via modern quality diversity algorithms (MAP-Elites, ME (Iso+LineDD), CMA-ME). The project is derived from the MarioGAN project. We use a newer version of the Mario-AI-Framework that allows for a richer definition of game tiles.
The GAN that generates Mario levels can be run by the following command in the GANTrain folder:
python3 GANTraining.py --cuda
However, training the GAN is unnecessary as we include a pretrained model in the repo.
Experiments can be run with the command:
python3 search/run_search.py -w 1 -c search/config/experiment/experiment.tml
The w parameter specifies a worker id which specifies which trial to run from a given experiment file. This allows for parallel execution of all trials on a high-performance cluster.