Score in maze environment

[CeHao-NUS]

Hello, I have a question about how the scores in the maze environments are calculated.

In the maze2d branch, we can get the reward and the normalized score.
The score is about 1.xx, but in the paper the score are 100.x. So did we just the original score to times 100?

Also, the variance is very small, even in the multi tasks, the variance is less than 1. I cannot reproduce these results. Can you help explain how to calculate the score in the maze environments? Thanks.

[sigmundhh]

Hi! This issue might be relevant. You're almost correct; they report the normalized accumulated reward times 100. This is obtained by:

• Summing up reward over the episode
• Getting the normalized accumulated reward (also called score) with env.get_normalized_score(total_reward) (see here)
• When reading the results with scripts/read_results.py they multiply the score from each run by 100 before averaging

Hope this helps!