Intelligent Level Generation for Super Mario

抄録

The computer game industry is growing rapidly with the yearly revenue raising up to tens of billion dollars. The market is huge but meanwhile very competitive. A computer game with static and deterministic design will be discarded after a short playing duration. Thus, we need a game that can keep creating interesting and challenging contents by adapting to customers’ preferences. In this research we aim to design an intelligent level generator for the famous Super Mario game. The generator collects the player’s data in a base game level and identifies his/her skill and featured tendencies such as jumping and coin-collecting. Then, it generates a customized game level by fitting to these tendencies. The level generator is directly evaluated by human players, and the parameters of the level generator are optimized through interactive evolutionary computation. A simple regression model is built and applied to make the evolutionary process more efficient and effective. Experimental results confirm the effectiveness of the optimization. The proposed level generator was also recognized as one of the best generators in the 2012 Mario AI Championship competition.