Developing a 2048 Player with Backward Temporal Coherence Learning and Restart

The puzzle game 2048 is a single-player stochastic game played on a 4 x 4 grid. It is very popular among similar slide-and-merge games. After the appearance of the game, several researchers developed computer players for 2048 based on reinforcement learning methods with N-tuple networks. The state-of-the-art player developed by Jaskowski is based on several techniques as the title of the paper implies. In this paper, we show that backward learning is very useful for 2048, since the game has quite a long sequence of moves in a single play. We also show a restart strategy to improve the learning by focusing on the later stage of the game. The learned player achieved better average scores than the existing players with the same set of N-tuple networks.