Guaranteeing the probability of success using repeated runs of genetic algorithm

Though genetic algorithm (GA) has found widespread application, there appears to be no guarantee of success or quantitative measure of the probability of success in a given application. This paper addresses this problem using the notion of repeatedly applying a GA. Several alternative interpretations of the algorithm are offered. The Q factor is introduced to characterize the efficacy of any GA. The repeated algorithm is applied to a six-degree object detection problem and experimental results are reported. A general methodology is given on the design of GA in a particular problem based on defining the maximum variation of a problem, using the training set to estimate the average probability of a single run to the desired level of statistical confidence, and using the testing set to verify the required performance. This paper paves the way for applying the GA to robust industrial applications for which the probability of convergence to the globally correct solution is required to be arbitrarily high. (C) 2001 Elsevier Science B.V. All rights reserved.