Hi SCOPE-RL team,
I am following the Real-World Dataset Example
and noticed the description of a few logged dataset fields:
step_per_trajectory: number of steps in a trajectory, int
n_trajectories: number of trajectories, int (optional)
size: number of data tuples, which is given by the multiplication of n_trajectories and step_per_trajectory, int (optional)
For real-world datasets, the number of steps per trajectory is often not fixed. This raises some questions:
Should step_per_trajectory be treated as a list of integers, where each entry corresponds to the length of one trajectory, rather than a single int?
If trajectories have variable lengths, then size != n_trajectories * step_per_trajectory in general. How should this be handled for compatibility with the OPE estimators in SCOPE-RL?
Is there an official recommendation or updated specification for preparing logged datasets with variable-length trajectories?
Clarification on these points would help users properly format real-world logged datasets for offline policy evaluation.
Thank you!
Hi SCOPE-RL team,
I am following the Real-World Dataset Example
and noticed the description of a few logged dataset fields:
step_per_trajectory: number of steps in a trajectory, int
n_trajectories: number of trajectories, int (optional)
size: number of data tuples, which is given by the multiplication of n_trajectories and step_per_trajectory, int (optional)
For real-world datasets, the number of steps per trajectory is often not fixed. This raises some questions:
Should step_per_trajectory be treated as a list of integers, where each entry corresponds to the length of one trajectory, rather than a single int?
If trajectories have variable lengths, then size != n_trajectories * step_per_trajectory in general. How should this be handled for compatibility with the OPE estimators in SCOPE-RL?
Is there an official recommendation or updated specification for preparing logged datasets with variable-length trajectories?
Clarification on these points would help users properly format real-world logged datasets for offline policy evaluation.
Thank you!