In this homework, you need to upload your averaged rewards after 1000 episodes. Because Kaggle requires at least two data rows, one for public and one for private data, you need to output the same score twice. An example "rewards.csv" is shown below:
Id
Predicted
FrozenLake8x8_public
0.312
FrozenLake8x8_private
0.312
We use Absolute Error (AE) as metric and set the target as 1. So if you get a score 0.312, you AE score will be 1 - 0.312 = 0.688. Lower AE means better model. The system will calculate the AE for you, so you should just update the averaged reward and don't reverse it.