Creatures That Cheat Their Creators
This newsletter is a bit shorter, but it contains a gem: one of my favorite papers featuring adorable but mischievous creatures. (I’m chilling on my couch with Covid, so I gotta take it a bit slow this week.)
In 2008, researcher Peter Krcah evolved some digital creatures so they could jump high (among other things). But instead of programming the creatures directly, they were placed in a digital environment and subject to (digital) evolution. Or to be more specific, subject to evolution was the body of the creatures and the control over the body (“brain”). Of course with neural networks. So from generation to generation the bodies and control changed, based on selective pressure: creatures that jumped higher were more likely to survive and reproduce.
It mostly worked, and he got his creatures to jump 7cm off the ground, with the cute creatures being around 15cm tall.
But that wasn’t always the end product when running his simulated evolution. Sometimes the evolution drifted towards taller bodies, so they would just reach higher. Why jump 10 cm when you can jump 1cm while being 9cm taller?
Clearly, a mistake in the fitness function. The fitness function measured the maximum height of the center of the mass. And so the creatures evolved to have this center high up.
This loophole could be fixed by measuring the jump height at the lowest point of the body. Or so it seemed. If it weren’t for the sneaky evolution. Why jump when you can fall in a creative way?
Because again, digital evolution is full of surprises: Watch how the creatures evolved.
Moral of the story: optimization can be full of surprises and is not always aligned with the creator of the optimization problem. There are many lessons here to be learned for machine learning and AI as well.
The jumping-creatures-anecdote is just 1 of the 32 entertaining anecdotes in the paper:
The stories are entertaining as much as they are educational. I think you will enjoy the paper as well.
Have fun!