Evolutionary Computation - Computational Intelligence in Time Series Forecasting - page 193

optimization problems, valuable in mathematics, natural sciences, and engineering.
The earliest attempts to map Darwin's ideas on to real-life problems was made by
John Holland and David Goldberg, who modeled many such problems. They
developed classifier systems , which are the predecessors of evolutionary systems .
Thereafter, accelerated work on evolutionary methods across the world was started.
The algorithms developed under the common term of evolutionary computation
generally start with the selection of an initial population as a possible initial set of
the problem solution. This is followed by stepwise iterative changing of the
selected population by random selection and use - in each iteration step - of
evolutionary operators like crossover, recombination, selection, and mutation in
order to improve the fitness of initial individual population members. Although
simple in principle, the evolutionary concept of computation has proven to be very
efficient in solving complex application problems that are not easily solvable using
traditional mathematical approaches.
In the meantime, depending on the nature of the problem to be solved, adequate
evolutionary algorithms have been developed, such as
x genetic algorithms (Holland, 1975), related to direct modelling of genetic
evolutionary processes
x genetic programming (Koza, 1992 and 1994), an extension of genetic
algorithms in which the population individuals are replaced by programs
x evolutionary strategies (1973), which model the evolution of evolution by
tuning the strategic parameters that control the changes in the evolutionary
process
x evolutionary programming (Fogel et al. , 1966), based on simulation of
adaptive behaviour of the evolution process
x differential evolution (Storn and Price 1995, 1996), a population-based
search strategy for optimizing real-valued, multi-modal objective functions.
As shown in this chapter, evolutionary algorithms are a special category of random
search algorithms . In contrast to traditional search algorithms like gradient
methods, which become impractical with the growing size of the search space,
evolutionary algorithms, because they are based on the population concept and are
operating with the genetic terms and operators, retain more or less the same size of
population over the generations and remain mathematically well manageable.
5.2 Genetic Algorithms
Genetic algorithms (GAs) are gradient free, parallel, robust search and
optimization techniques based on the laws of natural selection and genetics. The
GAs have confirmed their application power in solving practical problems which
are generally ill-defined, complex, and with multimodal objective functions. This
optimization technique is similar to its associated algorithms, such as simulated
annealing and other guided random techniques. GAs employ random search
algorithms aimed at directed location of the global optimum of the solution. The
algorithms are superior to the “gradient descent” methods that are not immune