Using Strongly Solved Mini2048 to Analyze Players with N-tuple Networks

2048 is a stochastic single-player game and there have been many studies of computer players for 2048. The authors believe that 2048 and its players can be useful for analyzing, comparing, and characterizing AI techniques. Yamashita et al. (2022) showed that a smaller variant of 2048, called Mini2048, can be strongly solved, and used the game to analyze some properties of AlphaZero-based players. In this study, we continue the work to deepen the analysis of the properties of the game itself and of the players. We first reproduce the retrograde analysis and then use the results to investigate the properties of the game, including the usefulness of the tile downgrading technique. We also develop four N-tuple networks following some of the state-of-the-art training methods and try to analyze the properties of N-tuple networks. We find several similarities and dissimilarities among the N-tuple networks. This is an important first step towards deep analysis of various AI techniques applied to 2048.