Question about the Design of the `i`th Agent's Context at `t`th Round
BBQGOD opened this issue · comments
I would like to thank you for the great work you've put into this project.
I've been exploring the codebase and I came across a particular design choice that intrigued me. I noticed that the context for the i
th agent at the t
th round is determined by the replies from agents 0-(i-1)
at the t
th round, along with the replies from agents (i+1)-(n-1)
at the t-1
th round.
I was under the impression that a more common approach might be to use replies from agents 0-(i-1)&(i+1)-(n-1)
at the t-1
th round. Could you possibly explain the rationale/intuition behind this design?
Hi,
Thanks for your interest in the project. Yes the latter of using replies from agents 0-(i-1)&(i+1)-(n-1)
at the t-1
round is more natural -- the newest version of the repo should be using that design choice.
Using the responses from 0-(i-1)
at the t
th round, along with the replies from agents (i+1)-(n-1)
at the t-1
th round was an unintended design choice in our original codebase (this design choice does encourage a more "debate-like behavior" where each agent responds to the new responses from other agents and might be an interesting direction of further exploration) -- I ran additional experiments and found that just using responses from all agents at round t-1
led to better results.