Specification network attributes |
Population size:
Sets the number of individuals in each generation.
Minimum weight:
The minimum value each weight can assume.
Maximum weight:
The maximum value each weight can assume.
Mutation rate:
During mutation one or more allele1 is randomly chosen and its value is changed.
The mutation rate is the probability of each gene2 being mutated each generation.
Generations:
This value sets the number of iterations that the GA will go through.
Terminating Error:
A floating point number greater than or equal to zero.
This is the error value at which you want the training of the network to cease.
Fitness Normalisation:
Check this box to include fitness normalisation.
With fitness normalisation the fitness value of each individual is normalised.
Elitism:
Check this box to turn on elitism.
This is a replacement strategy that takes the best individual and copies it into the next generation to ensure its features are not lost.
Choosing the number of cross over points |
Crossover is when two chromosomes3 join at one or more points and exchange genes2.
It is possible to specify the number of crossover points:
One point:
Chromosomes join at only on locus
Uniform:
Crossover at each locus is randomly determined.
n - point:
Chromosomes join at n loci.
Choosing the selection strategy |
Selection involves selecting the parents of the next generation.
There are many methods in existence, all of which are based upon the fitness of the individual.
Here you can choose from two strategies:
Roulette Wheel:
Each individual is given a slice of a virtual roulette wheel with each slice being proportional to fitness.
The wheel is spun once to select the individual.
Tournament:
Two individuals are selected via Roulette Wheel selection and their fitnesses are compared. The individual with the highest fitness value is selected.
Training with a GA |
Press the Train button to commence training.
Training will terminate when either the training RMS error drops below the terminating error threshold, or when the maximum number of generations is reached.
It may be the case that the training error ends up stopping at a lower value that the terminating error. This is because the error is evaluated at the end of each generation. Therefore at the end of one generation the error value may be above the terminating error, but by the end of the next generation the error value may have dropped well below the terminating error and this is the value that is displayed when training stops.
[1] Allele: [2] Gene: [3] Chromosome:This page is maintained by Melanie Middlemiss
mmiddlemiss@infoscience.otago.ac.nz
Last modified on: 22/12/98.