Mutation probability, genetic algorithms

Energy Barriers of Refolding. Thus every mutation operation of the GA started with a disruption of some stems in a given structure (an individual of the genetic algorithm population). Then after calculation of a list of stems compatible with the remaining structure, new stems were added by stepwise selection from the list, using a probability distribution for stem formation. The procedure continued until no stem could be added. 

For each benchmark in Tab. 2 we randomly picked 100,000 def/use classes as the training set, and parametrized the genetic algorithm with a population size of 100 individuals, 400 optimization generations, and a mutation probability of 10%. The values for these parameters were picked from experiences in early evaluation rounds we will analyze their impact on the approach more in detail in future work. 

Figure 7.1. Example of variablity in convergence of the genetic algorithm caused by random influences. The algorithm was run repeatedly ten times with an identical comhination of values of the adjustable parameters (population size = 28, probability of any modification = 90%, probability ratio qualitative mutation crossover = 10, asymptotic probability ratio quantitative mutation any modification = 0.75, coefficient of quantitative mutation = 10), starting from the first generation used for the population size equal 28.

Figure 7.5. Example illustrating that the diversity of the catalysts proposed by the genetic algorithm decreases more with increasing asymptotic probability ratio of quantitative mutation to any modification, as well as with increasing coefficient of quantitative mutation. The values of the remaining heuristic parameters were population size = 56, probability of any modification = 90%, probability ratio qualitative mutation crossover = 1.

---