Forward Modeling for Partial Observation Strategy Games - A StarCraft Defogger

Gabriel Synnaeve, Zeming Lin, Jonas Gehring, Vasil Khalidov, Nicolas Carion, Nicolas Usunier

Abstract: This paper we present a defogger, a model that learns to predict future hidden information from partial observations. We formulate this model in the context of forward modeling and leverage spatial and sequential constraints and correlations via convolutional neural networks and long short-term memory networks, respectively. We evaluate our approach on a large dataset of human games of StarCraft: Brood War, a real-time strategy video game. Our models consistently beat strong rule-based baselines and qualitatively produce sensible future game states.

TL;DR: This paper presents a defogger, a model that learns to predict future hidden information from partial observations, applied to a StarCraft dataset.

Keywords: forward modeling, partially observable, deep learning, strategy game, real-time strategy

Code: [ facebookresearch/starcraft_defogger](https://github.com/facebookresearch/starcraft_defogger)