Re:Genetic Algorithm 2D Car Thingy
by sklogic on Friday November 20, 2015 @09:45AM
(#10601364)
I played a bit with the "more faithful" genetic algorithms, got some promising results (getting out of the local minima better). The key is to use hierarchical encoding in your chromosome: instead of directly encoding traits/functions/properties/whatever, encode a program (maybe even in a Turing-complete representation) which in turn generates those traits/functions/properties/whatever, or even generates a next level of code. This is closer to how the biological systems work, and allows for much higher sensitivity to small changes.
[ Parent
| on HN ]
Do you have more information about this approach? I've never heard about it until now.
I tried to dig out the references from my mess of bibliography (I worked on this topic long before Mendeley and alike appeared), unsuccessfully.
I remember that I ran into this technique independently first (in a context of encoding expression trees) and then found a number of other examples in the literature (e.g., encoding the neural network topology, etc.).