You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I observed some problematic episodes from RunHistory files while training DRRN models on many tasks for simplification_str="". I am attaching handful episodes from Task20 task20_problematic_episodes.txt
There are many instances when the episodes terminate prematurely (isComplete=1) with a score of -1, without a logical reason.
They are accompanied by a java.lang.StackOverflowError message in the freelook, inventory, observation entries.
This seems to happen in runs when with simplification_str="", but not when with simplification_str="easy".
The text was updated successfully, but these errors were encountered:
Thanks for this report. This may be due to an issue with ScienceWorld -- it was tested extensively with the "easy" simplification flag, but it's possible that a randomly exploring agent like the DRRN might have uncovered a bug that's not present in the "easy" version of the game. I will put it on my TODO list to investigate this -- if you have any more reports from other tasks, please attach them. Right now it looks like it's limited to a small selection of cases where the electrical connection action is being used?
I observed some problematic episodes from RunHistory files while training DRRN models on many tasks for simplification_str="". I am attaching handful episodes from Task20
task20_problematic_episodes.txt
There are many instances when the episodes terminate prematurely (isComplete=1) with a score of -1, without a logical reason.
They are accompanied by a java.lang.StackOverflowError message in the freelook, inventory, observation entries.
This seems to happen in runs when with simplification_str="", but not when with simplification_str="easy".
The text was updated successfully, but these errors were encountered: