Systems need noise to be able to adapt.
Imagine a bullseye that lots of archers are trying to hit.
Each time an arrow connects, it generates a little burst of light.
It's natural for a team to try to optimize the accuracy of their archers.
This pull towards more efficiency is the most obvious thing in the world.
But now imagine the lights turn out, everything is totally dark.
For a while, the archers continue hitting the bullseye, and when they do, they see the light.
But then, all of a sudden, the light disappears–the arrows aren't connecting.
Unbeknownst to you, the target has moved.
How do you find it again?
You have to probe in the dark, sending arrows randomly to try to find a hit.
If the target continues moving, you might never find it.
If you would have had some noise in the arrows, some spread around the bullseye the chance is that one of the arrows would have kept hitting.
That would have shone the way for the other archers to update their aim.
This noise fundamentally allows sensing in the dark.
The "roving bullseye in the dark" is what actual targets are like in real environments.
A formal analysis I've seen has shown that the optimal amount of noise is proportional to the expected rate of movement of the target.
It's easy to forget in real life that the target is actually roving in the darkness, but you must never forget.
The "bullseye" that we can see is not the real target, it is a proxy for it.
It makes us forget that the bullseye we see can be misleading.