1

I have been experimenting with evolving artificial creatures, but so far all creatures just die. To initialize the creatures that do not result from asexual reproduction; I create around 8 random neurons which both have a connection in and a connection out. I'm using mutation to get a set of weights which are used in a small neural network, that can form recurrent connections. I have 15 inputs and 5 output. There is a max number of 25 neurons in the hidden layer. The mutation chance is 25%. The different mutations are add a connection, disable a connection, make a small change to a weight, add a neuron, and disable a neuron. Is there something off with my mutation chances?

Sam
  • 19
  • 3
  • Generally speaking a mutation rate of `25%` is pretty high, though your mile-age might vary depending on actual genetic problem at hand. Your problem description is a bit too vague: how are individuals "dying"? what is driving the evolutionary pressure on your population? how large is your population? how are individuals evaluated? what are the criteria for successful reproduction? what is the end-goal of your search? – Patrick Trentin Mar 12 '18 at 11:51

1 Answers1

0

Real evolution is a massively parallel computation. Even so it took eons to get the basics of life. And then most of them died. Only a small sliver of all possible genes are ok.

To get your simulation to work in a reasonable time frame, you're gonna have to take some shortcuts.

Also, you should make sure your "small neural net" is capable of creating the kind of lifeforms that are successful. Your architecture may not be powerful enough to produce viable life.

Ray
  • 2,974
  • 20
  • 26
  • Since this post I have gotten results. They aren't that great, but they work. After a few generations they all die out. How large of an artitechture do you think would be required to produce good results. – Sam Mar 05 '18 at 15:33