Toward Universal and Interpretable World Models for Open-ended Learning Agents

TL;DR

This paper presents a new class of interpretable, compositional Bayesian network-based world models designed for open-ended learning agents, enabling scalable, adaptive, and developmental learning through active model refinement.

Contribution

It introduces a generic, sparse Bayesian network class supporting interpretability and scalability, integrating structure learning with model-based planning for open-ended agents.

Findings

Supports broad stochastic process approximation

Enables active, developmental learning

Improves robustness and adaptability

Abstract

We introduce a generic, compositional and interpretable class of generative world models that supports open-ended learning agents. This is a sparse class of Bayesian networks capable of approximating a broad range of stochastic processes, which provide agents with the ability to learn world models in a manner that may be both interpretable and computationally scalable. This approach integrating Bayesian structure learning and intrinsically motivated (model-based) planning enables agents to actively develop and refine their world models, which may lead to developmental learning and more robust, adaptive behavior.