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

TL;DR

This paper introduces Finite Automata Extraction (FAE), a neuro-symbolic method that learns precise, generalizable world models from gameplay video using a novel domain-specific language, improving over neural network-based models.

Contribution

The paper presents FAE, a new approach that extracts finite automata-based world models from gameplay videos using a novel DSL, enhancing precision and generality over prior methods.

Findings

FAE learns more precise environment models than neural network approaches.

FAE produces more general code compared to previous DSL-based methods.

The approach enables better transfer and explainability of environment dynamics.

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.