There's lots new in the "let's evolve a fly" game I posted a few days ago - at that time someone pointed out that the maze was too difficult to navigate after even 2,000 generations.

So I simplified the maze. Now it's a race between various algorithms from Machine Learning. In the race are: purely evolved flies (a.k.a. crossover = parent flies' brains are mixed together!), flies that use backpropagation, flies that use transformer architecture (is attention really all you need?), and hybrids that use both an evolved algorithm and backpropagation. All this is unsupervised learning - they try to navigate on their own and learn from the results!

No spoilers, see for yourself if any of the evolved/hybrid flies can navigate the maze by clicking the maze button!

It's fun to pause it and click any of the flies' brains, you can see their whole neural net! Pause it at the beginning before any training and look at their brain, then look again after you've evolved it for a while!

This is ridiculously fun, and educational too. Click show statistics to see how the generations evolve over time. What's your high score?