Finite Automata Extraction: Low-data World Model Learning as Programs from Gameplay Video
This addresses the challenge of explainability and transfer in world models for AI systems, particularly in gaming or simulation domains, though it appears incremental as it builds on prior DSL-based methods.
The paper tackles the problem of learning explainable and transferable world models from gameplay video by proposing Finite Automata Extraction (FAE), which learns a neuro-symbolic model represented as programs in a novel domain-specific language called Retro Coder, resulting in more precise environment models and more general code compared to prior approaches.
World models are defined as a compressed spatial and temporal learned representation of an environment. The learned representation is typically a neural network, making transfer of the learned environment dynamics and explainability a challenge. In this paper, we propose an approach, Finite Automata Extraction (FAE), that learns a neuro-symbolic world model from gameplay video represented as programs in a novel domain-specific language (DSL): Retro Coder. Compared to prior world model approaches, FAE learns a more precise model of the environment and more general code than prior DSL-based approaches.