Deepracer Student League - What is happening when model training is "Evaluating"

Question

Hello everyone,

I was wondering what is happening during the "Evaluating" phase during training for deepracer student. I see through the simulation video that across evaluation runs, there are still big differences in deepracer performance, so I was wondering if it is comparing different policies that it formed during training. I am using the PPO algorithm. Thanks in advance!

Answer

[Here](https://docs.aws.amazon.com/deepracer/latest/developerguide/deepracer-console-train-evaluate-models.html#deepracer-examine-training-progress) is documentation regarding the reward graph and how to interpret it.

In AWS DeepRacer, training happens in iterations. Each iteration is collection of  episodes, n is 20 by default for PPO, and is configurable and 1 for SAC, and is fixed. At the end of every iteration, latest model is saved as checkpoint for evaluation. Evaluation runs for 5 episodes (called Trials) and evaluation metrics (average completion percentage, average reward value) are saved. The current selection for criteria best model is at the maximum average completion percentage.

Hope this information is helpful for you.

Deepracer Student League - What is happening when model training is "Evaluating"

Contenuto pertinente