Finite Automata Extraction: Low-data World Model Learning as Programs from Gameplay Video

World models are often neural network-based models that attempt to approximate an entire video game. Existing world models are not practical for game developers due to their large scale, lack of accessibility, their prompt-based interfaces, and their nature as black box models where developers cannot access the code. In this paper, we propose an approach, Finite Automata Extraction (FAE), that learns a neuro-symbolic world model represented as programs in a novel domain-specific language (DSL): Retro Coder. The neuro-symbolic world model structure allows for low data learning, meaning we can learn these models from a single gameplay video. This model’s structure means that developers can directly access code to change the behaviour of the model, and it’s low data cost makes it more accessible. In a comparison to prior neural world model approaches, FAE learns a more precise model of the environment and it learns more general code than prior DSL-based approaches.