Enhancing Artificial Intelligence Systems through ALife Principles and Generative Models: A Case Study with Super Mario Bros

GAN, DCGAN, Minimal VAE and MarioGPT can be located in the 'models' folder. Each folder includes the main scripts, such as:

• vae.py, vae_train, generate_levels_vae.py

• gan.py, gan_train.py, generate_levels_gan.py

• dcgan.py, dcgan_train.py, generate_levels_dcgan.py

A random latent vector, CMAES, a simple genetic algorithm, or MAP-Elites, can be used to generate levels.

Each model has at least one pre-trained option available and can be found under trained_vae, trained_gan, and trained_dcgan. To utilise the MarioGPT minimal implementation, please use:

• minimal_training_gpt.py
• minimal_generation_gpt.py
• simulation_gpt.py

Note:

• MarioGPT was re-trained, and there were no changes to the architecture, only slight variations in the training settings. The minimal_generation_gpt.py file offers the option to utilise prompts, such as prompts = ["many pipes, many enemies, some blocks, high elevation"].
• The pre-trained MarioGPT model can be found here (send a request for access)
• A sample of Gonzalez's original minimal VAE can be located at models/VAE/trained_vae/example.pt.

Virtual Environment and Installing Requirements

Create a virtual environment and use at least Python version >=3.9. Then, install the necessary requirements by running this:

pip install -r requirements.txt

Visualising Model

To visualise the latent space representation of a trained model, use:

python visualize.py

Run Simulation

To execute the simulation using 'simulator.jar', a Java version above 8 (at least OpenJDK 15.0.2) is required.

Play Level

To play a level yourself, use human_player=True, otherwise human_player=False. This will use Robin Baumgarten's A* agent instead. Ensure that visualize=True is also set to true to view the simulation.

python load_run_level.py

Bibliography and Software

Implementations are based, adapted and/or inspired by the following works:

Sudhakaran, s., Glanois, C., Freiberger, M., Najarro, E., & Risi S., 2023. MarioGPT: Open-Ended Text2Level Generation through Large Language Models. Available from: https://github.com/shyamsn97/mario-gpt/tree/main [Accessed 04 April 2023]

Paper: Sudhakaran, s., Glanois, C., Freiberger, M., Najarro, E., & Risi S., 2023. MarioGPT: Open-Ended Text2Level Generation through Large Language Models. ArXiv [Online]. Available from: https://arxiv.org/abs/2302.05981 [Accessed 04 April 2023].

Code basis based on and adapted from:

González-Duque, M., 2023. Minimal implementation of a Variational Autoencoder on Super Mario Bros (0.1). Available from: https://github.com/miguelgondu/minimal_VAE_on_Mario [Accessed 04 April 2023]

González-Duque, M., Palm, R. B., Hauberg, S., & Risi, S., 2022. Mario plays on a manifold: Generating functional content in latent space through differential geometry. In 2022 IEEE Conference on Games (CoG) (pp. 385-392). IEEE.

DCGAN architecture inspired by:

Volz, V., Schrum, J., Liu, J., Lucas, S. M., Smith, A., & Risi, S., 2018. Evolving mario levels in the latent space of a deep convolutional generative adversarial network. In GECCO. Aguirre, ed. Proceedings of the genetic and evolutionary computation conference, July 15 - 19, 2018, Kyoto Japan. New York US: Association for Computing Machinery, pp. 221-228.

Student-t approach inspired by:

Mathieu, E., Rainforth, T., Siddharth, N., & Teh, Y. W., 2019. Disentangling disentanglement in variational autoencoders. In International conference on machine learning (pp. 4402-4412). PMLR.

Map-Elites based and adapted from:

Bryon Tjanaka and Sam Sommerer and Nikitas Klapsis and Matthew C. Fontaine and Stefanos Nikolaidis, 2021. Using CMA-ME to Land a Lunar Lander Like a Space Shuttle. Available from: https://docs.pyribs.org/en/stable/tutorials/lunar_lander.html [Accessed 07.07.2023]

VGLC:

Summerville, A. J., Snodgrass, S., Mateas, M., Onta'n Villar, S., 2016. The VGLC: The Video Game Level Corpus Proceedings of the 7th Workshop on Procedural Content Generation

Mario-AI-Framework:

Created by Ahmed Khalifa, based on work by Sergey Karakovskiy, Noor Shaker, and Julian Togelius, which in turn was based on Infinite Mario Bros by Markus Persson. Available from: https://github.com/amidos2006/Mario-AI-Framework [Accessed 18.03.2023]