Bots and Simulations > DNA - General

Neural nets and mutations

(1/3) > >>

Moonfisher:
Is there a way to set point mutations to affect values far more often than opperators ?
I'm trying to evolve a neural network with point mutations, but it keeps breaking down the structure.
For instance if it produces no shell the output for shell could mutate it's store to an inc, and this would have a more noticable effect.
This would help the new bot survive but would also destroy the network leading to that output and prevent the chances for a more balanced shell management to evolve through the network.
I'd basicaly like to make the point mutations only affect values and not operators, if possible, or atleast balance it to something like 1 change of an operator for every 10000 value mutations.
It would also be usefull for evolving league bots to tweek their constants.
So if you're going to be messing with mutations anyway  

Testlund:
There's a slider on the mutations dialog where you can set the percentage for type and value for both Point Mutations and Copy error, if that's what you mean.

Numsgil:
That wouldn't help prevent store turning in to an inc, though.

Moonfisher:
Yeah I tryed fidling with that slider, but it doesn't seem to make any difference.
Problem is it seems that values and operators are regarded as the same thing, but in a neural network the values are the thickness of a connection, but if an operator is changes it'll most often correspond to a broken connection...
So when evolving a neural network you'll break pieces in the network about as fast as the weights are mutated, so theres litle to no chance for propper values to evolve before the network is completely destroyed by the mutations.

Numsgil:
As far as the code is concerned, a store is a type StoreCommand of value 1.  Whereas a number is of type Number with value 56 (or whatever).  So when you say only to change values, to the program that just it won't change a 56 in to a store.

Navigation

[0] Message Index

[#] Next page