search

weightagnostic / weightagnostic.github.io

repo for interactive article
https://arxiv.org/abs/1906.04358
Creative Commons Attribution 4.0 International
456 stars 110 forks source link

Time to generation 128 for biped and car driving? #12

Open lf94 opened 5 years ago

lf94 commented 5 years ago

I'm no expert in machine learning, probably couldn't even consider myself a beginner, but this paper really made me think: wow, surely this significantly reduces the time to train something reasonably well?

I would love to be able to train on a single high end video card (or even more amazingly... On a single CPU) within a reasonable time. I just want to know if we're approaching this possibility.

Thank you and I hope we can see some code soon! Or even better: how someone could build it from the ground up (ok, maybe too ambitious...)!

UltronBrain commented 5 years ago

可以先去看看neat进化算法，在强化学习方面效果很好，这个论文也是从neat上获得的灵感，我跑过neat的代码，只用了cpu，都没有用gpu，最后会出来一张网络的生成图，很便于观察。

lf94 commented 5 years ago

Thanks @nuistComputerScience . I'll check it out :slightly_smiling_face:

What kind of CPU, and how much time?

weightagnostic commented 5 years ago

Hi, the code is now released. We made it so that at least for the cartpole swingup experiment it runs on a single laptop.