Evolving Mario to Maximize Coin Score Using Neat and Novelty

Genetic algorithms can be used to evolve agents that will complete game tasks in a given game environment. In this paper, we discuss our experimental results using NEAT and Novelty to evolve Mario, from the popular game Super Mario Bros, to maximize his coin score. To conduct our experiments, we developed our own Mario simulator, creating a small world and a big world. Each world has an easy and hard version; the addition of coin boxes is what defines a world to be difficult. Just like in the original SMB, we added hidden rooms that contain many coins. However, finding the hidden room is the most difficult part. We hypothesized that Novelty would outperform NEAT because of its exploratory nature, while NEAT would focus more on getting the coins that are easy to obtain. Our results indicated that without an input that tells Mario what type of coin is nearest to him, Novelty significantly outperforms NEAT. However, with the input, NEAT can perform on par, or even better than Novelty at times. We argue that the reason for his discrepancy is that the addition of the input makes the task less deceptive for NEAT, which closes the performance gap between NEAT and Novelty.