Automatic Game Design

Today, my daily news network brought me to a blog post titled Automatic Game Design. The idea is novel and interesting from a researcher's perspective. It examines the theories of Ralph Koster and Juergen Schmidhuber, that one of the main contribution to the fun factor of games is the learnability factor. That is, fun games that those that are easy to learn but hard to master.

The theory is put to test by conducting an experiment with a game that is set up as a sequential series of possible moves an agent (player) can make. At the end of the game, the agent is given a score. Pacman is an example of a game that fits this scenario. The experiment pits a random agent that make random choices against a learning agent (eg, neural net, SVM, reinforcement learner) that tries to improve its score over time. If the learning agent consistently gets a better score compared to the random agent, the game is said to be learnable and therefore fun.

The idea is interesting, however neither the blog nor the paper has given any concrete evidence that learnable games areindeed fun games. For example, chess is a game that is eminentably learnable, but it is not necessarily a universally fun game for everyone. Thus, it ends up being a test of whether a game is learnable but there is much doubt as to whether said game is fun.