38
D. Morrall
GAs are a bottom-up or inductive technique that can, through dynamic
evolution, build the rules governing ecological systems. They offer a tool for
parameter optimization and for the development of structurally dynamic models.
Genetic algorithms can be used to solve a wide variety of problems but are most
commonly employed when: you don't know what set of instructions to give to the
computer to solve the problem (i.e., let the GA figure out the rules for you) the
dataset is very large and an exhaustive search for the solution is inefficient (e.g.,
traditional optimization techniques won't work) the data are fuzzy or there is
missing information the response surface is irregular (i.e., you need to find a
general solution) Koza (1992) goes as far as to say that genetic programming
provides "... a single, unified, domain dependent approach to the problem of
induction" (Figure 3.1).
3.3
Genetic Algorithm Design Details
Genetic Algorithms (GA) are computer solution-search and problem-solving
techniques based on the principles of evolution by natural selection. Through the
process of natural selection, GAs evolve linear, coded representations of data to
solve problems or develop strategies. Selection rules are designed by the
programmer to govern the direction that is taken to evolve solutions to problems.
Control strategies may be defined based on the programmer's conceptions about
how a system operates (e.g., select organisms that are better at procuring food) or
rules may be independent of internal system operation (e.g., optimize for
correlation between observed and predicted organism distributions). Rules may
be altered simply to produce a desired outcome (i.e., without any implied
causality) or to evaluate multiple hypotheses about how a system operates.
Although the concept of GA is simple, actual GA development involves
multiple design decisions and choices from among a wide variety of
implementation techniques (see Holland 1975 for a classic example).
The
optimization problem is coded as a finite-length string, often using a binary
representation. Therefore, it must be possible to represent the solution as this
finite-length string. Problem representation by a fixed-length string is the key
limitation to GAs. For more detail on GAs see Goldberg (1989). The basic
strategy includes the following procedures (Figure 3.2). GAs randomly create a
population of individuals. The mathematical representation of each individual
depends upon problem to be solved. Individuals in the population are evaluated to
determine their fitness (e.g., how well the individual produces the desired
outcome). The fitness evaluation is the most important determinant of how future
recombination is guided and the direction in which the population will evolve.
After fitness has been determined, individuals must be chosen to be parents.
There are a variety of methods that can be used for parent selection. Tournament
selection and roulette are 2 common methods. Once the parents have been
chosen, each parent is cloned to produce a child that is an exact replica. Genetic
material of the children is then exchanged via crossover or altered via mutation.
D. Morrall
GAs are a bottom-up or inductive technique that can, through dynamic
evolution, build the rules governing ecological systems. They offer a tool for
parameter optimization and for the development of structurally dynamic models.
Genetic algorithms can be used to solve a wide variety of problems but are most
commonly employed when: you don't know what set of instructions to give to the
computer to solve the problem (i.e., let the GA figure out the rules for you) the
dataset is very large and an exhaustive search for the solution is inefficient (e.g.,
traditional optimization techniques won't work) the data are fuzzy or there is
missing information the response surface is irregular (i.e., you need to find a
general solution) Koza (1992) goes as far as to say that genetic programming
provides "... a single, unified, domain dependent approach to the problem of
induction" (Figure 3.1).
3.3
Genetic Algorithm Design Details
Genetic Algorithms (GA) are computer solution-search and problem-solving
techniques based on the principles of evolution by natural selection. Through the
process of natural selection, GAs evolve linear, coded representations of data to
solve problems or develop strategies. Selection rules are designed by the
programmer to govern the direction that is taken to evolve solutions to problems.
Control strategies may be defined based on the programmer's conceptions about
how a system operates (e.g., select organisms that are better at procuring food) or
rules may be independent of internal system operation (e.g., optimize for
correlation between observed and predicted organism distributions). Rules may
be altered simply to produce a desired outcome (i.e., without any implied
causality) or to evaluate multiple hypotheses about how a system operates.
Although the concept of GA is simple, actual GA development involves
multiple design decisions and choices from among a wide variety of
implementation techniques (see Holland 1975 for a classic example).
The
optimization problem is coded as a finite-length string, often using a binary
representation. Therefore, it must be possible to represent the solution as this
finite-length string. Problem representation by a fixed-length string is the key
limitation to GAs. For more detail on GAs see Goldberg (1989). The basic
strategy includes the following procedures (Figure 3.2). GAs randomly create a
population of individuals. The mathematical representation of each individual
depends upon problem to be solved. Individuals in the population are evaluated to
determine their fitness (e.g., how well the individual produces the desired
outcome). The fitness evaluation is the most important determinant of how future
recombination is guided and the direction in which the population will evolve.
After fitness has been determined, individuals must be chosen to be parents.
There are a variety of methods that can be used for parent selection. Tournament
selection and roulette are 2 common methods. Once the parents have been
chosen, each parent is cloned to produce a child that is an exact replica. Genetic
material of the children is then exchanged via crossover or altered via mutation.