Classification Using an Online Genetic Algorithm

Genetic Algorithms (GAs) purport to mimic the behavior of natural selection. Many GAs, however, try to optimize their populations by means of a static fitness function — one that is derived from performance on a fixed set of examples. We propose an architecture for an online genetic algorithm (OLGA) for classification. An OLGA differs from standard genetic algorithms in that it does not repeatedly evaluate individuals against a fixed set of training examples. Instead, it is presented with a series of training examples, one at a time, and does not retain the entire set for training. Being online and incremental, OLGAs, like evolution strategies (Dasgupta & Michalewicz 1997), are applicable to tasks that require continuous learning to handle concept drift such as in adaptive systems as well as tasks in which the dataset is too large to be kept on hand for repeated evaluation (as in many online and interactive problems). By evaluating individuals on recent examples, OLGAs also better mimic the behavior of natural selection, as real organisms live in environments that are not identical to that of their ancestors. An OLGA is a GA, complete with a population of size p, fitness function, recombination operators, and mutation operators. However, unlike a traditional GA, the fitness of an individual changes over time, as it is exposed to more examples. In order to do so, we track the number of examples seen and the number of examples classified correctly for each individual and class. The fitness is calculated by summing over all classes the weight of the class times a function of correctcount and totalseen. The key idea in creating such a fitness function is to support newly created individuals so that they are not replaced before they have seen a reasonable number of examples (and thus have some estimate of their true fitness). The individuals can represent entire solutions, or can form a solution as a group, as is done in most classifier systems (Goldberg 1989). When an individual is not a complete solution (such as rule that fires only when it applies), we track the number of times the rule was applied (instead of the number of examples seen). The overall classification is by majority vote of those individuals making a classification. We have an implementation of an OLGA that we have applied to a set of related sample problems to test the feasibility of the architecture. The examples are a set of 6 boolean