A Genetic Algorithm for Data Reduction

When large amounts of data are available, choosing the variables for inclusion in model building can be problematic. In this analysis, a subset of variables was required from a larger set. This subset was to be used in a later cluster analysis with the aim of extracting dimensions of human flourishing. A genetic algorithm (GA), written in SAS®, was used to select the subset of variables from a larger set in terms of their association with the dependent variable life satisfaction. Life satisfaction was selected as a proxy for an as yet undefined quantity, human flourishing. The data were divided into subject areas (for example health, environment). The GA was applied separately to each subject area to ensure adequate representation from each in the future analysis when defining the human flourishing dimensions. GENETIC ALGORITHMS – A BRIEF INTRODUCTION Genetic Algorithms are iterative, heuristic (experience based) search processes that can be used to find solutions to problems where an exhaustive search of all potential solutions would be impractical due to time or resource constraints. John Holland (1993) was credited with their invention in the early 1970s. They mimic natural evolution by using techniques such as natural selection, inheritance, mutation and crossover. They are used in fields and contexts where the optimal solution is required from within a large search space. This research attempts to find the best subset of variables to summarise a larger dataset. Typically, a Genetic Algorithm (GA) starts with an initial population. This initial population consists of chromosomes that are traditionally a string of zeros and ones, and is usually a randomly generated sample of the solution search space. For example, for a variable selection / reduction exercise where there are 30 variables to choose from, each chromosome will be 30 bits long. Each of these bits is called an allele and each of these 30 alleles will consist of either a randomly generated 0 (indicating the variable is not to be selected) or 1 (indicating the variable is to be selected). Figure 1 is a visual example of two chromosomes, each 30 alleles long, from a population encoded to represent variable selection or nonselection from a total of 30 variables. Figure 1 Example chromosomes