Comment by jodrellblank
3 days ago
That is an interesting video! Your writing made me expect the neural network had changed gravity as hacky part of its solution, but it's the creator trying to train a neural network to balance a simulated double pendulum and failed, and then wondered if he could train a simpler version in a simulation with no gravity and high friction to make it easy, then see if the solution can gradually be retrained to cope as he adjusts those to full gravity and normal friction. The visualisations in it are very good.
Gah I can see why you would be confused
Yes yes, I was trying to show "here's a simplification one made to train a simple network how to do a hard thing" and much like removing pedals, I think it's effective
Unrelated, in dog agility they sometimes train weave poles using the channel method. So they set the weave poles to form a channel that the dog runs through. And then they gradually narrow the channel. Finally they have narrowed the channel until the poles are in a straight line