Speed of training

#1

Perhaps it’s because I’m used to the OpenAI Gym environments which are more simple than this environment, but training is slow-going; I seem to be able to achieve around 9K training episodes per day on a Ryzen 7 CPU. Is that similar to what others are getting? Is there any way to speed it up…?

#2

There’s an integrator_accuracy parameter in the environment constructor, which determines the accuracy of the simulation, a smaller value being more accurate and more time consuming. I believe the default is 0.00005, and I’ve been able to successfully set it up to about 0.005 (the environment crashes with larger values). Example: env=L2M2019Env(integrator_accuracy=0.005). It can make a noticeable difference in training speed, but nothing insane.

2 Likes
#3

Many thanks, I’ll give it a try