Domain-Independent Game Abstraction via Word Embedding

A recent research study introduces a versatile method for game abstraction that utilizes word embedding techniques derived from natural language processing. In this approach, actions are viewed as words, and gameplay data is considered a corpus, allowing for the training of word vectors that represent actions as real-valued vectors suitable for clustering and abstraction. The authors investigate fundamental embedding models, revealing that action embeddings can embody unexpected characteristics. The goal of this research is to extend game abstraction beyond the domain-specific contexts, such as poker, that have largely influenced previous studies.