Week 1: Natural Evolutional Strategy training best practice about practical_rl HOT 1 CLOSED

Well, unfortunately the best practice is to run on as many cores as you can find. As we discussed in the lectures, this family of algorithms achieves simplicity and scalability at the cost of poor sample efficiency.

If you don't have a machine with that many cores, try azure notebooks or google colab for a temporary source.
They both give out pretty good hardware for free, the catch is that they don't let you use it for long:

• Colab lasts for exactly 12 hours since it was created and restricts the jupyter interface a bit. The simplest way is to use !console commands.
• Azure can run indefinitely as long as you keep it active. You can write a selenium bot that sends some signal (e.g. save) every 2-3 minutes to make it more comfortable. Just don't tell 'em i was the one who gave you that idea.

from practical_rl.