Integrating active structure selection into factor-graph message passing

Integrate active selection strategies for generative model structure—specifically, Expected Free Energy–guided selection among candidate generative model structures with targeted data acquisition—into factor-graph message passing frameworks for variational free energy minimization, so that structure learning operates within the same local message-passing process.

Background

The paper discusses extending active inference beyond parameter learning to active selection of generative model structures, where Expected Free Energy (EFE) guides both choosing among candidate model structures and actively gathering data that best disambiguate those candidates.

Throughout the paper, inference and control are realized via variational free energy minimization implemented as message passing on factor graphs. The authors note that while EFE-driven active selection is conceptually formulated, it is not yet clear how to embed these strategies within the existing factor-graph message passing (e.g., CBFE-based, reactive message passing) framework.