EVALUATING PROGRAMS FORMED BY EXAMPLE: AN INFORMATIONAL HEURISTIC

Abstract

The construction of a procedure from a few examples of its execution--a major requirement for practical programming-by-example--is inevitably a drastically underdetermined problem. At the core of any incremental programming-by-example scheme is a heuristic module that creates and modifies the program, or "model," as it is formed. In general there are countless different ways that the model might plausibly be modified; the problem is to prune the set of candidate models to keep it down to a reasonable size. We develop a principled method for evaluating and comparing alternative models of a sequence of actions. Models are finite-state automata and therefore can contain branches and loops. Based on information theory, the method computes the entropy of a model in conjunction with the entropy of the sequence used to form it--a novel form of the "minimum description length" principle. The idea is to measure a model's predictive power, taking into account the extent to which it is justified by the sequence that has been used to create it. The performance of the measure is illustrated on test cases and accords with intuition about when sufficient evidence has accumulated to prefer a more complex model to a simpler one.