Procedural Content Generation of Rhythm Games Using Deep Learning Methods

The rhythm game is a type of video game which is popular to many people. But the game contents (required action and its timing) of rhythm game are usually hand-crafted by human designers. In this research, we proposed an automatic generation method to generate game contents from the music file of the famous rhythm game "OSU!" 4k mode. Generally, the supervised learning method is used to generate such game contents. In this research some new methods are purposed, one is called "fuzzy label" method, which shows better performance on our training data. Another is to use the new model C-BLSTM. On our test data, we improved the F-Score of timestamp prediction from 0.8159 to 0.8430. Also, it was confirmed through experiments that human players could feel the generated beatmap is more natural than previous research.