Has anyone tried to train the dodgeball example enrionment yet?
I dont get nearly the results that they have posted in the blog post for this project.
They have an elo of 2k on setp 40M while mine are still at around ~1200.
Any ideas what the reason could be?
One possibility could be that you’re using a different number of Unity instances to train. See this section:
“Result Variation Using Concurrent Unity Instances - If you keep all the hyperparameters the same, but change --num-envs=, the results and model would likely change.”
Yes that might be one reason. I used the config file provided in the repository which had env set to 3, but maybe they did have a different setup for the shown results.
