Modeling Race Track Difficulty in Racing Games
Summary
Recent years have shown a rising popularity of procedurally generated content, such as automated level design. To ensure the player enjoys the content, game developers need to make sure it is suitably difficult. This is challenging if at all possible when the content is generated after the product has shipped. The designers need to make sure the game can automatically alter the variables that control the difficulty, depending on the performance of the player. To determine which variables used in level generation control difficulty, a difficulty model is required. We attempt to find such a model for the racing game genre.
To identify what parts of the track most define the difficulty, we use two approaches. First, a data driven model, which uses machine learning to recognize difficult sections on the track. Second, an analytical model that attempts to predict where cars are most likely to lose traction, following the rules of physics. Using a custom-made racing game, our methods are tested empirically through player testing on various procedurally generated racetracks.
Results show that while we can not perfectly predict all difficult sections of a racetrack, crashes can indeed be predicted with above-average accuracy (over 60%) using simple algorithms, with relatively sparse data. The varying level of player performance is identified as one of the most influential reasons why accurate predictions are very hard to achieve. Further analysis of the data suggests some increased accuracy may potentially be achieved with slightly altered approaches.
Our exploratory work helps game developers identify at least the most problematic sections of tracks. We also believe it can be used as a foundation upon which further work can be based.