The article discusses a computational model designed to mimic how humans infer hidden rules by conducting experiments. The model is based on the principles that even if a rule is deterministic, the learner considers a broader space of fuzzy probabilistic rules, which are represented in natural language. The hypotheses are updated online after each experiment following approximately Bayesian principles. The model also incorporates experiment design according to information-theoretic criteria. The authors use the game Zendo, where players try to discover a hidden logical rule, as a context for their model. The model’s core principles include representing hypotheses in an inner symbolic language, updating beliefs in approximately Bayesian ways, and proposing experiments that are optimal in an information-theoretic sense.
Publication date: 12 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.06025
