GAINS: Comparison of Genetic AlgorIthms and Neural AlgorithmS for Video Game Playing.

In this paper we place ourselves in the context of sub-symbolic Artificial Intelligence. We aim at comparing two well known methods of learning (Neural Networks and Genetic Algorithms) for video game playing. The setting of video game playing (here we chose Super Mario Bros) is of particular interest because of the challenges it brings in terms of data collection. The data is challenging in nature due to its size (in our case the small number of levels in the game - thus fundamentally different from big data approaches) and its heterogeneity (in our case the different levels used to simulate non deterministic games). The non determinism aspect is key because we demonstrate it to be the main root cause of performance decline.