A limited survey of the effects of single parametric variation on the GENEtic search implementation system (GENESIS) package

The genetic algorithm is a heuristic that can be used effectively in the optimization of functions. Briefly, the algorithm selects good solutions out of a set as determined by an objective function. This set of solutions tends towards the optimal value of the objective functionmuch like natural selection where a population of strong animals lend to survive more than a population of weak animals. GENESIS is a public domain package authored by John Grefenstetle and released by the Navy Center for Research in Applied Artificial Intelligence. In this paper, we implement this package on the maximization of a function similar to that defined in Peng & Reggia 1987 (I and II) for optimal covering of symptom-diagnosis sets. We vary some chosen parameters of GENESIS one-at-a-lime to get a feel for the parameter's effect on the overall performance of the algorithm on this problem. Although this survey is by no means complete, we feel that this information may be useful in serving as guidelines, for anyone who wishes to implement the genetic algorithm on any problem. The parameters that we vary are: number of trials, population size, crossover rale, mutation probability, and the seed that is used by the random number generator. We present about 70 variations in these parameters for each of our 9 different experiments. Our control data were derived from an exhaustive search of the solution space. In our first set of variations, we attempted to find the relationship between the variation of each parameter and optimality in relatively small steps. Our second set of variations found the behavior of the data when these parameters were varied greatly. Through about 600 variation experiments (each containing 1024 solutions), we have found that the most influencial parameter for our problem, when varied singly, is the mutation rate. Through our wide-variation experiments we have also found the approximate settings where these parameters, when varied singly, will show decline in performance for our problem.