1: \begin{abstract}

2: We develop abstract learning frameworks  for synthesis that embody the principles

3: of the CEGIS (counterexample-guided inductive synthesis) algorithms in current literature.

4: Our framework is based on iterative learning

5: from a hypothesis space that captures synthesized objects, using counterexamples from an

6: abstract sample space, and a concept space

7: that abstractly defines the semantics of synthesis. We show that a variety of

8: synthesis algorithms in current literature can be embedded in this general framework.

9: %While studying these embeddings, we also generalize some of the synthesis problems these

10: %instances are of, resulting in new ways of looking at synthesis problems using learning.

11: We also exhibit three general recipes for convergent synthesis: the first two recipes based on

12: finite spaces and Occam learners generalize all techniques of convergence used in existing engines,

13: while the third, involving well-founded quasi-orderings, is  new,

14: and we instantiate it to concrete synthesis problems. 

15: \end{abstract}

16: