Idea: Evolve strategies for solving particular problems
Potential problems: Regression task Image classification
Next steps: Add a learning circuit Overall neural architecture: s = stim reception a = stim response (action) n = selection only response l = learned response u = (under selection) update circuit r = reward/death (can treat this as the same function for now, but should really be different, but highly correlated)
OUTPUT
a u l r
=====================
s || | x | | |
- - - - - - - - - - -
n || x | | | |
- - - - - - - - - - -
INPUT u || | | x | | - - - - - - - - - - - l || x | | | | - - - - - - - - - - - r || | x | | | - - - - - - - - - - - a || | | | x |
Learning
{s -> (n, l) -> a -> r} runs
{(s, r) -> u} runs
output of u assigned to l
(Repeat m times)?, then go through Selection
Selection
{s -> (n, l) -> a -> r} runs
Organism lives or dies based on r
Return to Learning
All neural circuits are under selection.
Additionally, the value of l is assigned during learning by the learning circuit
Future Ideas: Structure NN connections based on genes Instead of single neurons/connections, may want the ability for genes to specify entire common structures (including learning, IO, etc) with a small number of genes Sexual selection Place sexual metrics under natural selection as well