Evolving transformation sequences using genetic algorithms

Program transformation is useful in a number of applications including program comprehension, reverse engineering and compiler optimization. In all these applications, transformation algorithms are constructed by hand for each different transformation goal. Loosely speaking, a transformation algorithm defines a sequence of transformation steps to apply to a given program. It is notoriously hard to find good transformation sequences automatically, and so much (costly) human intervention is required. This work shows how search-based meta-heuristic algorithms can be used to automate, or partly automate the problem of finding good transformation sequences. In this case, the goal of transformation is to reduce program size, but the approach is sufficiently general that it can be used to optimize any source-code level metric. The search techniques used are random search (RS), hill climbing (HC) and genetic algorithms (GA). The paper reports the result of initial experiments on small synthetic program transformation problems. The results are encouraging. They indicate that the genetic algorithm performs significantly better than either hill climbing or random search