2
$\begingroup$

After witnessing the rise of deep learning as automatic feature/pattern recognition over classic machine learning techniques, I had an insight that the more you automate at each level, the better the results, and I, therefore, turned my focus to neuroevolution.

I have been reading neuroevolution publications with the same desire to automate at every level.

Do genetic algorithms evolve? Do they get better at searching through the solution space for each generation over time? Is this legitimately "evolution"?

$\endgroup$
  • $\begingroup$ This question reminds me very much of ai.stackexchange.com/q/8854/3217 ... Maybe I should have a look at Wikipedia if it doesn't become clear from the relevant pages there. $\endgroup$ – Martin Thoma Nov 20 '18 at 6:48
2
$\begingroup$

A genetic algorithm is a class of evolutionary algorithms.

They do get better at searching through the solution possibilities for each trial (generation) over time because evolution usually starts from a population of randomly generated individuals, and is an iterative process. In each generation, the fitness of every individual in the population is evaluated. The more fit individuals are stochastically selected from the current population, and each individual's genome is modified to form a new generation thus getting better and better with each generation over time.

Evolution is defined as the gradual development of something, especially from a simple to a more complex form.

$\endgroup$
  • 1
    $\begingroup$ I think the OP originally meant if the actual algorithms, like the mutation algorithm, also evolves or not, not if genetic algorithms are a form of evolution. $\endgroup$ – nbro Nov 17 '18 at 23:17
1
$\begingroup$

The answer of Steve Okay is correct, but I'm not sure if it really answers the question to the necessary degree.

It is important to focus on the first part:

A genetic algorithm is a class of [...] algorithms.

That means they act in boundaries that can be described exactly. It's only about searching good parameters / combinations.

Differences to evolution by natural selection are:

  • Fitness function: in natural selection, this is nature itself. On the most basic level the laws of physics, on higher abstraction levels chemistry and biology. For the algorithms, it is whatever the developer thinks fits to the problem. It might be a bad choice that steers development in the wrong direction
  • Sample sizes: think of how many bacteria there were until the first mammal developed. You will not even be close to the order of magnitude in a computer program
  • Breeder selection: see fitness function
  • Mutation: see fitness function; also, the amount of possible mutations is way bigger than in the algorithms.
  • Generations/time: similar as with the sample size; nature has a head start of a few billion years. And, after all, our computers are just one part of nature (aka our universe)

Now I would like to answer in a different direction: evolutionary algorithms are just one tribe of machine learning. And one which is not successful as well. The last publication I read from this branch was about 3 years ago so please correct me if something happened I'm not aware of. But at the time they couldn't even get MNIST to 90%, where a beginners tutorial on deep learning already gets to 95% with cheap hardware in less than one hour. So while they might be able to search through a lot of possible solutions, I'd rather invest the money you would spend on computational power in a deep learning working student. The student will be faster and cheaper.

$\endgroup$
-1
$\begingroup$
The Computer Scientist have developed optimization tool which is known as  Evolutionary Algorithm. It evolve a neural network with a code included, building the perfect deep learning network involves a hefty amount of art to accompany sound science.
An Evolutionary Algorithm uses the mechanisms inspired by the biological evolution, such as reproduction, mutation, recombination, and selection. 

Evolutionary algorithms often perform well all approximating solutions to all types of problems because they ideally do not make any assumption about the underlying fitness landscape.

$\endgroup$
  • $\begingroup$ You wrote "often". Please give at least 3 examples. $\endgroup$ – Martin Thoma Nov 20 '18 at 7:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.