Procedural content generation via ML

In the free time, I enjoy studying game development and I’m particularly intrigued by the use of AI techniques for the automated generation of content.

My focus is on the generation of functional (as opposed to cosmetic) content. Simply put, game levels or rules rather than textures or sound effects. The challenge with functional content is that it must satisfy very complex constraints.

For instance, unsurpassable obstacles like huge gaps or walls can make a platformer level impossible to complete. Contrarily to the generation of images, a single bit change could turn a playable level into an unplayable one.

As a joint research effort with multiple thesis students, we developed Constrained Adversarial Networks (paper), a neuro-symbolic deep generative model that was used, among other things, for the generation of playable Super Mario Bros. levels and stable molecules.