Evolving Controllers for Mario AI Using Grammar-based Genetic Programming

Abstract

Video games mimic real-world situations and they can be used as a benchmark to evaluate computational methods in solving different types of problems. Also, machine learning methods are used nowadays to improve the quality of non-player characters in order (i) to create human like behaviors, and (ii) to increase the hardness of the games. Genetic Programming (GP) has presented good results when evolving programs in general. One of the main advantage of GP is the availability of the source-code of its solutions, helping researchers to understand the decision-making process. Also, a formal grammar can be used in order to facilitate the generation of programs in more complex languages (such as Java, C, and Python). Here, we propose the use of Grammar-based Genetic Programming (GGP) to evolve controllers for Mario AI, a popular platform to test video game controllers which simulates the Nintendo's Super Mario Bros. Also, as GP provides the source-code of the solutions, we present and analyze the best program obtained. Finally, GGP is compared to other techniques from the literature and the results show that GGP find good controllers, specially with respect to the scores obtained on higher difficulty levels.