In the Decision Transformer implementation, I noticed that the return-to-go (RTG) is scaled in Gym environments (e.g., by dividing by a constant like 1000), but in Atari environments, it is not scaled at all.
Why is this difference in scaling applied?
In the Decision Transformer implementation, I noticed that the return-to-go (RTG) is scaled in Gym environments (e.g., by dividing by a constant like 1000), but in Atari environments, it is not scaled at all.
Why is this difference in scaling applied?