Selection (evolutionary algorithm)

Selection is a genetic operator in an evolutionary algorithm, a metaheuristic inspired by biological evolution that aims to solve hard problems at least approximately. It has a dual purpose: it chooses individual genomes from a population for breeding, for example with the crossover operator, and it also chooses which candidate solutions carry into the next generation. The biological model is natural selection. The basis for choosing is the quality of each individual, measured by the fitness function. Keeping the best individuals unchanged from one generation to the next is called elitism. The best known procedures include fitness proportionate or roulette-wheel selection, where the chance of being picked is proportional to fitness, and stochastic universal sampling, which spins one wheel with several equally spaced pointers. Others are rank selection, steady state selection, tournament selection (the best of a randomly chosen subset), truncation selection (a top share of the population), Boltzmann selection with a falling temperature, and lexicase selection, which judges individual training cases separately. The methods differ mainly in selection pressure: high pressure makes a population converge faster but can explore the search space too little, while low pressure may stop it converging at all.