question about train and development mode about slm-lab HOT 3 CLOSED

Hi, dev mode is simply train mode with rendering of environment and network parameter update check. Because these two operations are expensive, it will also run slower, and is used only for dev.

from slm-lab.

Hi, dev mode is simply train mode with rendering of environment and network parameter update check. Because these two operations are expensive, it will also run slower, and is used only for dev.

So doesn't RL involve the validation procedure? It will find the best policy during training procedure and we just use the policy on the test dataset to evaluate its generalization ability. Is it?

from slm-lab.

RL validation/eval is different from supervised learning. It is done online as the policy/network/agent iterates, so evaluation is ran at checkpoints at regular intervals. See https://github.com/kengz/SLM-Lab/blob/master/slm_lab/experiment/control.py#L69-L82
Also note that in RL the training data is also the test data. This is still an area of research.

from slm-lab.