Answers are grounded in this paper and cite their sources.
We study the learnability of classes of computable structures under models that allow finitely many mind changes. Extending classical notions of explanatory learning from informant and from text, we introduce new paradigms where the information source and the convergence requirements are modified. In particular, we define Δ 2 0 -learning, where each atomic fact may be presented with finitely many errors before stabilizing, and c.e.- and d.c.e.-learning , where information is restricted to positive atomic facts that may either never be retracted (c.e.) or be retracted at most once (d.c.e.). We provide syntactic characterizations for these notions: Δ 2 0 -learning coincides with definability by finite existential sentences, and c.e.-learning coincides with TxtEx-learning. For d.c.e.-learning we give partial characterization in restricted languages. Furthermore, for n -learning from informant, where learners are allowed at most n mind changes, we establish a complete syntactic characterization in terms of specific infinitary formulas of bounded depth.
No takes yet. Share an insight, caveat, or question.
Fokina et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: