查看文章

We address the problem of computing an abstraction for a set of examples, which is precise enough to separate them from a set of counterexamples. The challenge is to find an over-approximation of the positive examples that does not represent any negative example. Conjunctive abstractions (e.g., convex numerical domains) and limited disjunctive abstractions, are often insufficient, as even the best such abstraction might include negative examples. One way to improve precision is to consider a general disjunctive abstraction.

We present , a new algorithm for learning general disjunctive abstractions. Our algorithm is inspired by widely used machine-learning algorithms for obtaining a classifier from positive and negative examples. In contrast to these algorithms which cannot generalize from disjunctions, obtains a disjunctive abstraction that minimizes the number of disjunctions. The result …