Canonical EA Definition

The general scheme for an EA optimisation consists of a number of operators used in succession :

• Initialization Sets of parameters (genes) encoded as strings (chromosomes) are initialized, generally randomly, within their domain constraints.
• Evaluation The strings are then assessed using the evaluation function to give their fitness values. Often these fitness values are stored alongside the chromosomes themselves in some data structure.
• Selection Selection schemes determine which of the strings should be propagated into the next generation. These schemes often generate a probability of propagation for each string.
• Reproduction A new set of candidate strings are generated from the old ones. Some method is used to take two strings from the old population (parents) to produce two new strings (offspring).
• Mutation Elements of strings are changed randomly under some probabilistic selection scheme.
• Replacement Offspring replace their parents in the new population if they have a better fitness value.

A canonical EA is a subset of reproductive population algorithms. These are algorithms that keep a collection of candidate solutions that are iteratively improved over successive generations. The following characteristics are those that represent a large subset of those found in the literature:

• All candidate solution vectors are the same length.
• The population of solutions is always the same size as the algorithm progresses.
• Reproduction between populations always uses two parent chromosomes, chosen by some probabilistic selection strategy.
• Reproduction is performed by crossover operations.
• Mutation is possible, according to some predefined population-wide mutation probability.