Has anyone managed to retrain the ML soccer agents on their home machine? How long did it take?

[vag_liv]

Hi all,

I am trying to retrain the soccer agents from the ML agents soccer example.
I haven’t changed anything from what was supplied (trainer files, reward functions etc), this was just a test to see how long it would take on my home PC.

I have GPU acceleration, with a Geforce GTX 1080.

With PPO traner, I ve left it running all night, and the mean reward seemed to be stuck at 0.0 at the end.
With Sac trainer, i ve left it running for around 5 hours, to get similar results.(but maybe i wasnt meant to use Sac with this network, since the soccer example wasnt in the sac trainer yaml file, i added it by copying from the trainer.yaml)

Anyone can give me any idea of how long would training the soccer agents take? (to get to a quality similar to what the supplied training network from the Unity examples has)

[vincentgao88]

@vag_liv

Hi all,

I am trying to retrain the soccer agents from the ML agents soccer example.
I haven’t changed anything from what was supplied (trainer files, reward functions etc), this was just a test to see how long it would take on my home PC.

I have GPU acceleration, with a Geforce GTX 1080.

With PPO traner, I ve left it running all night, and the mean reward seemed to be stuck at 0.0 at the end.
With Sac trainer, i ve left it running for around 5 hours, to get similar results.(but maybe i wasnt meant to use Sac with this network, since the soccer example wasnt in the sac trainer yaml file, i added it by copying from the trainer.yaml)

Anyone can give me any idea of how long would training the soccer agents take? (to get to a quality similar to what the supplied training network from the Unity examples has)

Since Soccer environment is a adversarial environment, it is suppose to converge at 0 reward.