ucla-rlcourse / competitive-rl Goto Github PK

I have spent whole day investigating the huge performance gap when evaluating some users' agents in cloud (colab) and in local (local Mac or Linux).

Some agent achieves -90 reward in my environment but get +800 in cloud.

I have checked:

1. The model can be successfully restored.
2. The action computing code is identical.
3. The model responses identically to the same input.
4. The python packages are identical.

So the only potential reason for this is the observation. In the above figure, one can see that the only major differences is the font.

If font really making differences? I have invite LIU Lixin and LIU Hanyuan to test (Their agents' performance is bad). As one can see in the following figure, the font in original rendering pictures are really different. But the grassland colors are identical. So the conclusion is that the fonts in different platform have different behavior. Not every platform has the pygame.font.SysFont('Comic Sans MS', 5) font!

Please resolve this issue. @Edwardhk @1155107756