Semi-Supervised Tile Embeddings: A General, Multi-Game Level Representation

Representing video game levels for level generation and analysis tasks remains an open problem. Existing approaches generally rely on hand-authoring or are game-specific. Tile embeddings are a general machine learned-representation for tile-based game levels, however they have thus far relied solely upon hand-authored representations of levels for training data. In this paper, we introduce Semi-Supervised Tile Embeddings (SSTE) which make use of semi-supervised learning to allow for training on levels lacking human authored representations. We evaluate SSTE over many experiments, finding that it performs equivalently or better than existing tile embeddings. Thus SSTE stands as the first general machine-learned level representation that can scale without requiring additional human labor.