"Building Mario Levels with Machine Learning | AI and Games #39" is a video published by AI and Games that explores academic research in the Super Mario franchise. The video discusses recent and ongoing research in applying machine learning to Super Mario level generation, focusing on the transition from building levels that adopt player telemetry to mimicking the original designs from Mario titles. The video highlights four notable bodies of research in recent years, exploring a diverse range of AI methods and techniques employed in the process.
Dr. Steve Dahlskog's research at the University of Malmo focused on identifying design patterns within Mario levels and using evolutionary computation to build and assess levels based on the number of patterns found. This led to increased level variety and more accurate interpretations of Mario level design.
Adam Summerville, an assistant professor at California State Polytechnic University, explored two distinct approaches to Mario level generation using machine learning. The first approach used Markov chains and Monte Carlo tree search (MCTS) to validate the quality of generated levels. The second approach used long short-term memory (LSTM) networks to train against levels from the original Super Mario Brothers and The Lost Levels, generating new levels that shared similar properties while remaining novel.
Matthew Guzdial's research at Georgia Tech involved learning about Mario levels by watching people play them on YouTube. The system identified high interaction areas in the video footage and learned how sprites are positioned relative to one another in segments of video footage. This allowed the system to generate levels that were astoundingly accurate for a system that learned from video footage.
Dr. Vanessa Volz's research at the Technical University of Dortmund involved building levels using generative adversarial networks (GANs). The generator created solutions to a given problem while the discriminator evaluated their quality. The resulting levels were generated very quickly and could be adapted to specific design properties.
Overall, machine learning techniques have been shown to provide a viable option for procedurally generating Mario levels with a level of quality that surpasses much of the existing work in the field. There is great potential for these techniques to be employed in other detail-based games with their own unique level design principles.
