There are two competing factors that need to be balanced in the selection process, the selective pressure and genetic diversity. Selective pressure, the tendency to select only the best members of the current generation to propagate to the next, is required to direct the GA to an optimum. Genetic diversity, the maintenance of a diverse solution population, is also required to ensure that the solution space is adequately searched, especially in the earlier stages of the optimization process. Too much selective pressure can lower the genetic diversity so that the global optimum is overlooked and the GA converges to a local optimum. Yet, with too little selective pressure the GA may not converge to an optimum in a reasonable time. A proper balance between the selective pressure and genetic diversity must be maintained for the GA to converge in a reasonable time to a global optimum.

IML offers two variants of a standard technique for the selection process commonly known as tournament selection (Miller and Goldberg; 1995). In general, the tournament selection process randomly chooses a group of members from the current population, compares their objective values, and picks the one with the best objective value to be a parent for the next generation. Tournament selection was chosen for IML because it is one of the fastest selection methods, and offers you good control over the selection pressure. Other selection methods such as roulette and rank selection may be offered as options in the future.

In the first variant of tournament selection, you can control the selective pressure by specifying the tournament size, the number of members chosen to compete for parenthood in each tournament. This number should be two or greater, with two implying the weakest selection pressure. Tournament sizes from two to ten have been successfully applied to various GA optimizations, with sizes over four to five considered to represent strong selective pressure.

The second variant of tournament selection provides weaker selective pressure than the first variant just described. The tournament size is set at two, and the member with the best objective value is chosen with a probability that you specify. This best-player-wins probability can range from 0.5 to 1.0, with 1.0 implying that the best member is always chosen (equivalent to a conventional tournament of size two) and 0.5 implying an equal chance of either member being chosen (equivalent to pure random selection). Using this option, you could set the best-player-wins probability close to 0.5 in the initial stages of the optimization, and gradually increase it to strengthen the selective pressure as the optimization progresses, in a similar manner to the simulated annealing optimization technique.

Another important selection option supported in IML is the elite parameter. If an elite value of is specified, then the best solutions will be carried over to the next generation unchanged, with the rest of the new population filled in by tournament selection, crossover, and mutation. Setting the elite parameter to one or greater will therefore guarantee that the best solution is never lost through selection and propagation, which often improves the convergence of the algorithm.