A Further Investigation of Neural Network Players for Game 2048

Game 2048 is a stochastic single-player game. Development of strong computer players for Game 2048 has been based on N-tuple networks trained by reinforcement learning. Some computer players were developed with neural networks, but their performance was poor. In our previous work, we showed that we can develop better policy-network players by supervised learning. In this study, we further investigate neural-network players for Game 2048 in two aspects. Firstly, we focus on the component (i.e., layers) of the networks and achieve better performance in a similar setting. Secondly, we change input and/or output of the networks for better performance. The best neural-network player achieved average score 215 803 without search techniques, which is comparable to N-tuple-network players.