IB is only guaranteeed/proven for rewards in [0,1].
Currently, code silently accepts rewards outside this range.
We could either raise an error if this happens or (better) normalize any rewards on the fly as part of the ib machinery, so it can handle realistic envs that yield arbitrary rewards.
IB is only guaranteeed/proven for rewards in [0,1].
Currently, code silently accepts rewards outside this range.
We could either raise an error if this happens or (better) normalize any rewards on the fly as part of the ib machinery, so it can handle realistic envs that yield arbitrary rewards.