This paper presents an approach to determine the optimal Genetic Algorithm (GA), i.e. the most preferable type of genetic operators and their parameter settings, for a given problem. The basic idea is to consider the search for the best GA as an optimization problem and use another GA to solve it. As a consequence, a primary GA operates on a population of secondary GAs which in turn solve the problem in discussion. The paper describes how to encode the relevant information about GAs in gene strings and analyzes the impact of the individual genes on the results produced. The feasibility of the approach is demonstrated by presenting a parallel implementation on a multi-transputer system. Performance results for finding the best GA for the problem of optimal weight assignment in feedforward neural networks are presented.