Improving DNN-based 2048 Players with Global Embedding

2048 is a popular game for which plenty of computer players have been created. However, many created 2048 players, especially all DNN-based ones, only implicitly use tile values as inputs and access tile position information. In this study, we take one of the best DNN-based 2048 players as a baseline and propose a 2048 player directly using both tile values and tile positions as inputs. Additionally, we explore the possibility of embedding all tile values and positions that we then concatenate with the network's regular value inputs. We first train these variations in a short session and then select the best two models with the baseline to be further trained in a long session. Our best two methods performed better than the baseline DNN player in both short and long training sessions.