Abstract  Recent years have seen a growing interest in player  modeling for digital games. Goal recognition,  which aims to accurately recognize players’ goals  from observations of low-level player actions, is a  key problem in player modeling. However, player  goal recognition poses significant challenges  because of the inherent complexity and uncertainty  pervading gameplay. In this paper, we formulate  player goal recognition as a sequence labeling task  and introduce a goal recognition framework based  on long short-term memory (LSTM) networks.  Results show that LSTM-based goal recognition is  significantly more accurate than previous  state-of-the-art methods, including n-gram encoded  feedforward neural networks pre-trained with  stacked denoising autoencoders, as well as Markov  logic network-based models. Because of increased  goal recognition accuracy and the elimination of  labor-intensive feature engineering, LSTM-based  goal recognition provides an effective solution to a  central problem in player modeling for open-world digital games