Play BlueTeam
Investigating Agent Behavior In different RL methods's itch.io pageResults
| Criteria | Rank | Score* | Raw Score |
| Reproducibility | #12 | 2.449 | 3.000 |
| Generality | #14 | 1.633 | 2.000 |
| ML Safety | #14 | 1.633 | 2.000 |
| Mechanistic interpretability | #15 | 0.816 | 1.000 |
| Novelty | #15 | 0.816 | 1.000 |
Ranked from 2 ratings. Score is adjusted from raw score by the median number of ratings per game in the jam.
Judge feedback
Judge feedback is anonymous.
- It's an interesting project, but it mainly focuses on analyzing the specific training of different RL algorithms rather than delving into mechanistic interpretability. It would be intriguing to see the project extended to include analyses of how the models differ in their internal representation of the problem space, in addition to plotting the training dynamics. Additionally, providing more commentary on these topics would have given more context. Overall, it's a great starting point for interpretability research!
What are the full names of your participants?
Al-Hitawi Mohammed Abed , Saif Ali and Bertold Pal
What is your team name?
Blue Team
What is you and your team's career stage?
MSc students
Does anyone from your team want to work towards publishing this work later?
Yes
Where are you participating from?
Budapest
Leave a comment
Log in with itch.io to leave a comment.
Comments
In Deepmind's RL lectures there is also a fair comparison between several methods from a statistical inference point-of-view and less of a empirical one. (it is also probably backup up with their hands-on experience).
I would love a clarification on how you see it in connection to the hackathon.