Curriculum learning reward thresholding bug fix #1141
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
pderichai
force-pushed
the
develop-cl-bug-fix
branch
from
7efabd7 to
bedefab
Compare
awjuliani
reviewed
Sep 4, 2018
| if ((progress > self.data['thresholds'][self.lesson_num]) and | ||
| (self.lesson_length > self.data['min_lesson_length'])): | ||
| if progress > self.data['thresholds'][self.lesson_num]: | ||
| print(progress, 'is above the threshold, successfully incrementing lesson') |
Contributor
There was a problem hiding this comment.
Why do we print here instead of logging?
Contributor
Author
There was a problem hiding this comment.
Woops, that was a debug print. Let me remove that.
Contributor
Author
|
Merging, approved offline. |
Contributor
Author
|
Actually merging now, approved offline again. |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Curriculum thresholding based on rewards was shown to be broken #895. This PR redefines
min_lesson_lengthto be the minimum number of episodes that must be completed in a lesson. Once the minimum number of episodes has completed, the curriculum will become eligible for increment.Changes
reward_buffer. This buffer is a fixed-size queue that stores the cumulative reward given by the most recent episodes completed by the trainer.trainer_controller.pyuses the average cumulative reward in thereward_bufferto determine whether the reward threshold has been met.trainer_controller.pywill set the size of the capacity ofreward_bufferwhen constructing the PPOTrainer. The size of the reward buffer must be at least as large as themin_lesson_lengthin the curriculum, and in this implementation the size of the buffer is set to exactly that.lesson_lengthfield has been removed fromCurriculumsince it was a vague metric. The burden of ensuring that the minimum number of episodes have completed is ontrainer_controller.py.