Fitness Distance Correlation as a Measure of GA Performance

In this paper, the mathematical interpretation of correlation coefficient is reviewed to explain the conditions under which it operates. Using the work of Jones and Forrest (1995) on fitness distance correlation (FDC) as a measure of problem difficulty for genetic algorithms, a novel framework combing FDC with the Experimental Design perspective in statistics is proposed. It is shown that this method not only satisfies the mathematical condition of correlation coefficient, but alson that it is closely relevant to genetic operators such as crossover and mutation and can therefore be used to predict the performance of genetic algorithms more accurately. Different well-known problems such as epistasis interactions, isolation or needle-in-a-haystack, high fitness variance,deceptiveness and multimodality, which make the GA search process difficult, are investigated. Experimental results show that this framework is an effective metric for GA performance on the fitness landscape and offers useful guidance in constructing efficient genetic algorithms.