>>> you are still talking about a final optimization problem of 4^1,000,000,000.
There is no final optimization step that analyze the 4^1,000,000,000 possibilities. We are not the best possible human-like creature with 1,000,000,000 pairs of bases.
> method of gradient descent
Do you know the method of gradient descent? Nice. It is easier to explain the problem if you know it. In the method of gradient descent you don't analyze all the possible configurations and there is no guaranty that it finds the absolute minimum. It usually finds a local minimum and you get trapped there.
For this method you need to calculate the derivatives, analytically or numerically. And looking at the derivatives at a initial point, you select the direction to move for the next iteration.
An alternative method is to pick e few (10? 100?) random points nearby your initial point, calculate the function in each of them and select the one with the minimum value for the next iteration. It's not as efficient as method of gradient descent, but just by chance half of the random points should get a smaller value (unless you are to close to the minimum, or the function has something strange.)
So just this randomized method should find also the "nearest" local minimum.
The problem with the DNA is that it is a discrete problem, and the function is weird, a small change can be fatal of irrelevant. So it has no smooth function where you can apply the method of gradient descent, but you can still try picking random points and selecting one with a smaller value.
There is no simulation that picks the random points and calculate the fitness function. The real process in the offspring, the copies of the DNA have mutations and some mutations made kill the individual, some make nothing and some increase the chance to survive and reproduce.
Would you please not be a jerk on HN, no matter how right you are or how wrong or ignorant someone else is? You've done that repeatedly in this thread, and we ban accounts that carry on like this here.
If you know more than others, it would be great if you'd share some of what you know so the rest of us can learn something. If you don't want to do that or don't have time, that's cool too, but in that case please don't post. Putting others down helps no one.
who is talking about neurons? Beneficial random mutations propagate, negative don't, on average. In this way, the genetic code that survives mutates along the fitness gradient provided by the environment. The first self-propagating structure was tiny.
It's not literally the gradient descent algorithm as used in ml, because individual changes are random rather than chosen according to the extrapolated gradient, but the end result is the same.
>computationally intractable nature of 4^1,000,000,000
which is a completely wrong number, even if only because of codon degeneracy. Human dna only has 20 amino acids + 1 stop codon, which are encoded by 64 different sequences. Different sequences encode the same amino acid.
Again, this thread is about the computationally intractable nature of 4^1,000,000,000.
Got math? A proof maybe to support your statements?