Playing Game 2048 with Deep Convolutional Neural Networks Trained by Supervised Learning

Abstract

Game 2048 is a stochastic single-player game and 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 (convolutional) neural networks, but their performance was poor. In this study, we develop computer players for Game 2048 based on deep convolutional neural networks (DCNNs). We increment the number of convolution layers from two to nine, while keeping the number of weights almost the same. We train the DCNNs by applying supervised learning with a large number of play records from existing strong computer players. The best average score achieved is 93, 830 with five convolution layers, and the best maximum score achieved is 401, 912 with seven convolution layers. These results are better than existing neural-network-based players, while our DCNNs have less weights.