This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
Interpretability Hackathon
Accepted at the 
Interpretability Hackathon
 research sprint on 
November 15, 2022

Probing Conceptual Knowledge on Solved Games

We explored how a Deep RL agent uses human interpretable concepts to solve connect-four. Based on 'Acquisition of Chess Knowledge in AlphaZero' paper by DeepMind and Google Brain, we used TCAV to explore concepts detection in RL agent for connect four. Our agent architecture was inspired by AlphaZero and trained using the OpenSpiel library by DeepMind. Our novelty is in the decision to study connect four as it was solved with a knowledge based approach in 1988. Which means that to some extent we understand this game better than chess!

Amir Sarid, Bary Levy, Dan Barzilay, Edo Arad, Itay Yona, Joey Geralnik
4th place
3rd place
2nd place
1st place
 by peer review