Settings for the genetic algorithm model

Each generation creates an entirely new pool of models by selecting the fittest ones from the original pool as parents, and recombining them to produce new offspring.

Each generation replaces a certain number of models from the pool. In each generation, the fittest models from the original pool are selected as parents and recombined to produce new offspring. The new offspring replace the worst models in the original pool. This algorithm tends to converge faster than the generational algorithm.

Each generation uses every model as a parent. After randomly selecting another parent, the offspring are created by recombining the parents. The offspring replace the parents only if they are fitter than the parents. This ensures a monotonically increasing average fitness.

This algorithm uses mutation to create offspring. Each generation mutates every model to create new offspring. If the fitness of the offspring is better than their parent, they replace the parent. Otherwise, there is still the probability of acceptance determined by the Boltzmann equation (difference in fitness divided by the current temperature). After each generation, the temperature is decreased by using the specified decrease factor. The simulated annealing algorithm is designed to circumvent premature convergence at early stages.

If the best and average performance in a pool have not improved for several generations, try switching to this technique to produce new models and, after some time, select one of the other genetic algorithm techniques.

The relative fitness of a model, when compared to all other models in the pool, determines the probability of its selection as a parent. This is known as the stochastic universal version of the roulette wheel selection. The stochastic universal mechanism produces a selection that is more accurate in reflecting the relative fitness of the models than the steady state mechanism.

The relative fitness of a model, when compared to all other models in the pool, determines the probability of its selection as a parent. This is known as the roulette wheel selection because the process is similar to spinning a roulette wheel in which fitter models have more numbers on the wheel relatively to less fit models. There is a greater probability of selecting a highly fit model. The wheel is spun to select each parent.

This method randomly picks a certain number of models as contestants for a tournament. The fittest model in this collection wins the tournament and it is selected as parent.

The raw fitness values are used to determine the selection probabilities of models. However, this can also lead to premature convergence when some of the models have exceptionally high fitness values. Before using raw fitness values, rescale the fitness values by using an alternative scaling method.

Using this method, the fittest model is given a fitness s, between 1 and 2. The worst model is given a fitness of 2-s. Intermediate models get the fitness value given by the following interpolation formula:

f(i)=s-(2(i-1)(s-1))/(N-1), where i=1..N

Exponential ranking gives more chance to the worst models at the expense of those above average. The fittest model gets a fitness of 1.0, the second best is given a fitness of s (typically, about 0.99). The third best is assigned s^2 and so on. The last one receives s^(N-1).

Linear scaling adjusts the fitness values of all models in such a way that models with average fitness get a fixed number of expected offspring.

If the minimum yields a positive scaled value:

scaled f=f*(s-1)/(max-avg)+(max-s*avg)/(max-avg)

Otherwise:

scaled f=f*1/(avg-min)-min/(avg-min)

In both cases, the average always gets a scaled value of 1. In the first case, the maximum is assigned a scaled value of s, whereas, in the second case, the minimum is mapped to 0.

This scaling method introduces a moving baseline. The worst value observed in the most recent generations is subtracted from the fitness values, where s is known as the window size, typically between 2 and 10. This scaling method increases the chance of selecting the worst model, which prevents the pool from prematurely optimizing around the current best model.

This scaling method dynamically determines a baseline based on standard deviation. It sets the baseline s, and the standard deviation below the mean, where s is the scaling factor, typically between 2 and 5.