Hi, there! I’m learning Rllib for training agent along with a outside simulator. I found that the
PolicyServerInput is suitable for this. but I’m not sure where the
PolicyServerInput would be running when the
num_works > 0.
I looked into the codes, it seems the
PolicyServerInput running in remote rollout worker.
IIUC, in a distributed ray cluster,
PolicyServerInput would listen on different address, that would be not easy for the policyClient to connect, Because it doesn’t know the rollout worker’s address.
Do I make a mistake?