Chapter 5: Implementing a GA

Section 5.4: Selection Methods

For all of the techniques below, what are the advantages? What are the disadvantages?

1. Roulette Wheel.
2. Stochastic Universal Sampling (SUS). Spin a wheel once and make the selection with n evenly spaced pointers.
3. Sigma Scaling. The expected value of individual i at time t, Exp(i,t) = 1 + [ (f(i) - f_avg(t)) / 2 σ(t) ]
4. Elitism.
5. Boltzman. Exp(i,t) = e^f(i)/T / e^f_avg(t)/T. Decrease T over time.
6. Rank. One linear technique is to order the population from 1 to N based on increasing fitness.
   Exp(i,t) = Min + (Max - Min)[(rank(i,t) - 1)/(N-1)] where 1 ≤ Max ≤ 2 and Min = 2 - Max.
   Exponential techniques can also be used.
7. Tournament. Generate a random number, r, where 0 ≤ r ≤ 1. Choose two chromosomes randomly and if r > k (a constant) select the fitter parent, otherwise select the less fit parent.
8. Steady State. The generations overlap. Define the generation gap to be the fraction of new individuals in a generation.

Section 5.5: Genetic Operators

For the operators discussed below, when will they work well? When will they not work so well?

Crossover

• Single point
• Two point (Multipoint)
• Uniform - can be parameterized with p, often 0.5 ≤ p ≤ 0.8
• Many more...

Mutation

Others

• Inversion
• Crowding - the new individual replaces the most similar individual
• Fitness Sharing
• Mating Restrictions - spatial, similarity, non-similarity, tags, etc.

Section 5.6: Parameters for GAs

• p_m
• p_c
• N, the population size
• and many more ...

Dejong did a very elaborate experiment that showed that across many problems, good settings are p_m - 0.001, p_c = 0.6 and 50 ≤ N ≤ 100.

Some researchers have investigated how to use a GA to evolve these parameters. This is called a meta-GA.

Some researchers suggest that the parameters should adapt during the run. This is called self-adaptation.