Tutorials

The session introduced the General Video Game AI framework and the competition that grew around it. Agents in this setting must act without prior knowledge of the game, which makes planning, learning, and generality visible in the same problem.

The tutorial moved beyond the basic competition tracks to a newer feature: games described in VGDL could be parameterized, opening a space of variants around one set of rules. The presenters showed how an evolutionary method, the N tuple Bandit EA, could tune those parameters for balance or challenge.

Participants left with a concrete way to treat a game as a design space rather than a fixed test case.

This tutorial connected procedural content generation with machine learning. It began with the reasons designers generate content, then narrowed to methods that learn from existing game artifacts instead of relying on hand written rules.

The guided project used the Video Game Level Corpus, a collection of levels and maps from classic games. Participants worked through data representation, training, and sampling with multidimensional Markov chains and long short term memory networks. The session closed on evaluation and the open problems that come with small data sets.

No machine learning background was required, but the project portion asked for programming experience and a computer.