bcbcb/geneticalgorithms

#Genetic algorithms

Genetic algorithms take an initial population, select survivors, mutate them, and repeat the process until a goal state is reached or resources are exhausted.

##Steps For example, consider some problem where that involves selecting an ideal 9-character string, e.g. 'abcdefghi'.

####Initial generation To begin, we generate an initial generation of x=2 bloodlines, each with n=3 random offspring.

[['abcdefgai', 'abcdxfghi', 'wtrbekvkl'], ['helaworad', 'awfluehsk', 'seklfkese']]

####Select survivors via fitness function From each bloodline we select a survivor based on a fitness function. Our fitness function is related to the problem; for this arbitrary example, let's just make it return the number of 'a' characters in the string.

for each bloodline:
  for each offspring:
    apply the fitness function
  select the top offspring
['abcdefgai', 'helaworad']
return top survivor if the goal condition was met or if resources have been exhausted.

####Spawn and mutate offspring Now we create offspring from these survivors; we do this by copying and mutating each survivor n times. In our example, mutating will mean replacing one character with a random character.

for each survivor:
  create n offspring by copying the survivor and mutating that copy
these offspring are the new generation
[['zbcdefgai', 'abcdefgao', 'abcaefgai'], ['helawozad', 'heaaworad', 'hexaworad']]

####Repeat Repeat the select and spawn/mutate steps until either a goal is met or resources have been exhausted. A goal condition could be, for example, at least 5 'a' characters in the string. Resources could be CPU cycles, time, or some other consumable unit; perhaps 1000 would be an appropriate limit.

That's all there is to it!

##Notes Genetic algorithms are useful for converging on 'good not perfect' solutions more quickly than brute force solutions. If an obvious fitness function is available, that may be a sign that genetic algorithms are a viable approach.

Keep in mind that genetic algorithms are not guaranteed to (and most likely won't) find the perfect solution, so only use this approach when there are strong resource constraints or approximations are acceptable.