Intensive Crossovers: Improving Convergence and Quality in a Genetic Query Optimizer

Resumen. Database schemas and user queries are continuously growing with the need for storing and accessing large amounts of structured information. Among the several proposals to deal with the Large Join Query Problem, genetic optimizers have been shown to be a competitive approach. We propose a new search strategy to improve the quality and convergence of genetic query optimizers. We call our first technique Intensive Crossovers (IC) and it shows that, in terms of quality of the results, it is worthier to spend more time creating extra child plans locally in a crossover operation than to focus on crossover operations on a lot of different execution plans. After the first analysis of IC we propose an improved technique called Increasing Intensive Crossovers (IIC). The idea behind this improvement is to speed-up the convergence of IC. All in all, we show that the search strategy of choice is paramount to determine the convergence and quality of a genetic query optimizer. Our work opens a new line of research oriented to unlink genetic optimizers from their dependency on the random effects of, both, the initial population and the random decisions taken through the optimization process.