Interpreting Neural-Network Players for Game 2048

Abstract

Game 2048 is a stochastic single-player game and development of strong computer players for 2048 has been based on N-tuple networks trained by reinforcement learning. In our previous study, we developed computer players for game 2048 based on convolutional neural networks (CNNs), and showed by experiments that networks with three or more convolution layers performed much better than that with two convolution layers. In this study, we analyze the inner working of our CNNs (i.e. white box approach) to identify the reasons of the performance. Our analyses include visualization of filters in the first layers and backward trace of the networks for some specific game states. We report several findings about inner working of our CNNs for game 2048.