Rule Learning Through Active Inductive Inference

We propose a grammar-based approach to active inference based on hypothesis-driven rule learning where new hypotheses are generated on the fly. This contrasts with traditional approaches based on fixed hypothesis spaces and Bayesian model reduction. We apply these two contrasting approaches to an established active inference task and show that grammar-based agents' performance benefits from the explicit rule representation underpinning hypothesis generation. Our proposal is a synthesis of the active inference framework with language-of-thought models, which paves the way for computational-level descriptions of false inference based on an aberrant hypothesis-generating process.