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

2025/08/15 by Goel, Dave, Guzdial, Matthew, Sarkar, Anurag
#Artificial Intelligence (cs.AI) #FOS: Computer and information sciences

paper · doi:10.48550/arxiv.2508.11836

Abstract

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.

Citations

• Generative AI in Game Development: A Qualitative Research Synthesis
• Diffusion Models Are Real-Time Game Engines
• KnowPC: Knowledge-Driven Programmatic Reinforcement Learning for Zero-shot Coordination
• iVideoGPT: Interactive VideoGPTs are Scalable World Models
• Learning of Generalizable and Interpretable Knowledge in Grid-Based Reinforcement Learning Environments
• Mastering Diverse Domains through World Models
• MarioNette: Self-Supervised Sprite Learning
• Playable Video Generation
• Mastering Atari with Discrete World Models
• DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning
• Learning to Simulate Dynamic Environments with GameGAN
• Neural Game Engine: Accurate learning of generalizable forward models from pixels
• Analysis of Explainers of Black Box Deep Neural Networks for Computer Vision: A Survey
• Few-Shot Bayesian Imitation Learning with Logical Program Policies
• Recurrent World Models Facilitate Policy Evolution
• World Models
• RobustFill: Neural Program Learning under Noisy I/O
• DeepCoder: Learning to Write Programs

Related

• Human in the Loop: Interactive Passive Automata Learning via Evidence-Driven State-Merging Algorithms
• Extracting Finite Automata from RNNs Using State Merging
• A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning
• Disentangled Sequential Autoencoder
• Explaining Transition Systems through Program Induction