Professional Documents
Culture Documents
GA detailed process
Schema theory
2 GA details
the stages through which a GA passes in evolving the genome population: At Generation n:
1. Create a mating pool of the same size as the population by
i. applying the fitness function to every genome;
ii. randomly selecting genomes from the population, based on the biased roulette wheel;
iii. mutating genes in the mating pool (alternatively, mutation may occur at the step 2.iii, or not at
all).
2. Build a new population (Generation n+1) by:
i. selecting pairs of genomes in the mating pool for crossover (those not selected pass directly
into the new population);
ii. mutating genes on crossover (alternatively, mutation may occur at step 1.iii, or not at all).
3. Go back to step 1.
This process is repeated until very good solutions emerge.
3 GA details
Schema theory:
What are the values of o(H) and δ(H) for the schema
101**0*1?
ANSWER
H has five fixed positions, so its order, o(H) = 5. The first
fixed position is position 1 and the last is position 8, so its
defining length, δ(H) = 7.
17 Schema theory
Schema theory:
The basic idea behind schema theory is that fragments of the genome
(the schemata’s fixed positions) are responsible for contributions to
an individual’s overall fitness.
A single schema will be tested many times each generation, once for
each individual that contains it in the population.
Finding the fitness of an individual also contributes to the
information about the relative fitness of all the schemata that can
make up that individual.
The selection step of the GA will generally select those individuals
with the highest fitness and therefore those schemata that most
contribute to that fitness will be more commonly represented.
18 Schema theory
Schema theory:
For instance, consider the task of trying to find a genome that encodes for
the largest binary number.
Let’s take the number that a genome encodes as its fitness.
If we have a small population of six individuals, as shown in the table, we
can see that individuals with the 1***** schema have an average fitness of
45.7, while individuals with the 0***** schema have an average fitness of
20.3.
Similarly, the schema *****1 has a fitness of 31 and *****0 has a fitness
of 34.
There is a clear difference in fitness between 1***** and 0*****, but
*****1 and *****0 have very similar fitnesses.
Any sensible GA will be more likely to select the fittest individuals, so we
can expect that the next generation will have more instances of 1***** than
0*****, but there will be little pressure to select *****1 over *****0, so
the proportions of these schemata in the next generation will be more open.
19 Challenges of GA
Bootstrap problem:
In artificial evolution, the danger is that all the new
variants on an initial population will be of negative or zero
fitness, or at best of such low fitness that no improvement
is possible and the evolutionary process cannot get started.