Alchemist: Learning guarded affine functions

We present a technique and an accompanying tool that learns guarded affine functions. In our setting, a teacher starts with a guarded affine function and the learner learns this function using equivalence queries only. In each round, the teacher examines the current hypothesis of the learner and gives a counter-example in terms of an input-output pair where the hypothesis differs from the target function. The learner uses these input-output pairs to learn the guarded affine expression. This problem is relevant in synthesis domains where we are trying to synthesize guarded affine functions that have particular properties, provided we can build a teacher who can answer using such counter-examples.We implement our approach and show that our learner is effective in learning guarded affine expressions, and more effective than general-purpose synthesis techniques.