Mitchell, T. J. (2010) An exploration of evolutionary computation ...

3.2 Injecting Diversity
Several variations on the traditional EA model counterbalance the loss of diversity (leading
to convergence) with the continual introduction of novel genetic material.
One approach is to ensure that each offspring satisfies a measure of uniqueness before
being accepted into the population. Offspring that fail to meet the required criterion are
systematically mutated until they are sufficiently different from the rest of the population.
This technique was adopted by Whitley and Kauth (1988), and Mauldin (1984) to improve
the performance of the simple GA. Mauldin‘s GA applies a variable uniqueness
requirement which is decreased throughout the course of evolution; the assumption being
that diversity is most important in the early stages of evolution. Gradually reducing the
uniqueness level ensures the eventual convergence of the population at a single point. This
approach was found to improve the off-line performance of the GA.
Similar results may be achieved by adopting very high rates of mutation (Grefenstette,
1986). Cobb (1990) introduced a hypermutation system which comes into effect when it is
assumed that diversity is being lost. The traditional GA system is employed whilst fitness
is progressing, but when there is a measured decline in progress (population convergence),
the GA switches into a hypermutation mode (high mutation rate) to restore diversity.
The sudden introduction of new genetic material has a similar effect to the complete re-
initialisation of the population, a method examined in Krishnakumar (1989) and Mathias et
al (1998), termed cataclysmic mutation by Eshelman (1990). In other circumstances
mutation has been substituted for an entirely stochastic system, in which randomly
generated solutions are inserted directly into the population as evolution takes place
(Bonham and Parmee, 2004).
These injection approaches to diversity preservation have been criticised as addressing the
symptom of the problem rather than the cause. The question then arises: what are causes of
diversity loss in traditional EAs? In his PhD thesis Mahfoud (1995) extensively examined
the primary causes of diversity loss within the GA. Shir and Bäck (2005) later reconsidered
Mahfoud‘s observations from an ES perspective. There are three major factors that lead
traditional EAs towards suboptimal convergence:
33