Fitness Distance Correlation Analysis: An Instructive Counterexample

Fitness distance correlation (FDC) has been offered as a summary statistic with apparent success in predicting the performance of genetic algorithms for global optimization. Here, a counterexample to Hamming-distance based FDC is examined for what it reveals about how GAs work. The counterexample is a fitness function that is ``GA-easy'' for global optimization, but which shows no relationship between fitness and Hamming distance from the global optimum. Fitness is a function that declines with the number of switches between 0 and 1 along the bitstring. The test function is ``GA-easy,'' in that a GA using only single-point crossover can find the global optimum with a sample on the order of $10^{-3}$ to $10^{-9}$ of the points in the search space, an efficiency which increases with the size of the search space. This result confirms the suspicion that predictors for genetic algorithm performance are vulnerable if they are based on arbitrary properties of the search space, and not the actual dynamics of the genetic algorithm. The test function's solvability by a GA is accurately predicted, however, by another property---its evolvability, the probability that the genetic operator produces offspring that are fitter than their parents. It is also accurately predicted by FDC that uses not Hamming distance, but a distance measure defined by the crossover operator itself. A comparison is made between Hamming-distance based FDC analysis, crossover-distance based FDC analysis, evolvability analysis, and other methods of predicting GA performance.

[1]  Philippe Collard,et al.  Fitness Distance Correlation in a Dual Genetic Algorithm , 1996, ECAI.

[2]  Colin R. Reeves,et al.  An Experimental Design Perspective on Genetic Algorithms , 1994, FOGA.

[3]  Patrick D. Surry,et al.  Fitness Variance of Formae and Performance Prediction , 1994, FOGA.

[4]  John J. Grefenstette.  Predictive Models Using Fitness Distributions of Genetic Operators , 1994, FOGA.

[5]  Hans-Georg Beyer,et al.  Toward a Theory of Evolution Strategies: Some Asymptotical Results from the (1,+ )-Theory , 1993, Evolutionary Computation.

[6]  Ingo Rechenberg,et al.  Evolutionsstrategie : Optimierung technischer Systeme nach Prinzipien der biologischen Evolution , 1973 .

[7]  Tim Jones.  Evolutionary Algorithms, Fitness Landscapes and Search , 1995 .

[8]  Bernard Manderick,et al.  The Genetic Algorithm and the Structure of the Fitness Landscape , 1991, ICGA.

[9]  C. Cockerham,et al.  An Extension of the Concept of Partitioning Hereditary Variance for Analysis of Covariances among Relatives When Epistasis Is Present. , 1954, Genetics.

[10]  David B. Fogel,et al.  Using fitness distributions to design more efficient evolutionary computations , 1996, Proceedings of IEEE International Conference on Evolutionary Computation.

[11]  Lee Altenberg,et al.  The Schema Theorem and Price's Theorem , 1994, FOGA.

[12]  Gunar E. Liepins,et al.  Deceptiveness and Genetic Algorithm Dynamics , 1990, FOGA.

[13]  Terry Jones,et al.  Fitness Distance Correlation as a Measure of Problem Difficulty for Genetic Algorithms , 1995, ICGA.

[14]  Georges R. Harik,et al.  Foundations of Genetic Algorithms , 1997 .

[15]  Prügel-Bennett,et al.  Analysis of genetic algorithms using statistical mechanics. , 1994, Physical review letters.

[16]  Kalyanmoy Deb,et al.  Long Path Problems , 1994, PPSN.