Replies: 3 comments 8 replies
|
I'll be updating this as I continue testing. @ejunprung |
2 replies
|
Btw, here is the Product Delivery results. NativeRL 1.8.0 - https://app.pathmind.com/sharedExperiment/17265 1.8.0 works well but in 1.7.1, the reward is cleared out completely (i.e. all 0s). @slinlee Do you happen to know where the reward terms stuff is in the webapp code? |
6 replies
|
Update: I created a spreadsheet because editing the table above sucked. https://docs.google.com/spreadsheets/d/1LUpvfqk9yLMuoGngS9JZ86Oro3jSoC7bhqW3_XLRqTg/edit?usp=sharing |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
1.7.1 was pre-reward terms
1.8.0 introduced reward terms and reward balancing while maintaining compatibility with classic reward functions. Found to have worse performing policies in some cases
1.8.1 - enhances reward balancing and fixes some bugs.
1.8.1 is available on test.devpathmind.com, but right now we need to update the webapp to use the new 'use_auto_norm' flag
results are on this spreadsheet https://docs.google.com/spreadsheets/d/1LUpvfqk9yLMuoGngS9JZ86Oro3jSoC7bhqW3_XLRqTg/edit?usp=sharing
--- old info ---
1.7.1 - agv - classic reward function https://test.devpathmind.com/sharedExperiment/7049 the results look terrible. Throughput is 2. I think the webapp's format for setting up trainings isn't compatible with 1.7.1 anymore. (attn @ejunprung so we'll test locally??)
1.8.1 - agv - classic reward function https://test.devpathmind.com/experiment/7044
All reactions