Dormant program nodes and the efficiency of genetic programming

摘要

In genetic programming, there is a tendency for individuals in a population to accumulate fragments of code - often called introns - which are redundant in the fitness evaluation of those individuals. Crossover at the sites of certain classes of intron cannot produce a different fitness in the offspring, but the cost of identifying such sites may be high. We have therefore focused our attention on one particular class of non-contributory node that can be easily identified without sophisticated analysis. Experimentation shows that, for certain problem types, the presence of such dormant nodes can be extensive. We have therefore devised a technique that can use this information to reduce the number of fitness evaluations performed, leading to substantial savings in execution time without affecting the results obtained.