Bots don't know they've been stabbed

[belhilly]

In classic 7-player bot games, the bots sometimes don't realize I've stabbed them, and continue to act as though allied with me.

E.g. look at Turkey's behavior after I (Austria) stabbed them in 1909 in this game: https://webdiplomacy.net/board.php?game ... #gamePanel

I attacked them in Apulia in spring 1909, then they supported me in Denmark in the fall! They continued to avoid attacking me, even as I continued to take their centers.

[LeonWalras]

A large part of the training data was points per supply centre games, so the bots are more than happy to play for second place. It is my hope that a newer bot version will expunge these games now we have a larger pool of finished games, but I don't think anyone is volunteering to do this.

[belhilly]

I'm not sure that playing for second place explains it. Once I'm clearly hostile why don't they attack me? In the game I linked above, Turkey could have taken Warsaw and probably other centers from me if they wanted to -- instead they continued to treat me as an ally while I ate them up.

[A_Tin_Can]

I don't know that your solo was really stoppable at that point - a human might not fight back then either.

[A_Tin_Can]

Those particular bots do suffer from a lack of understanding of stalemate lines - so I don't think it was really coming from a place of realising that you were going to win and couldn't be stopped, though. Similarly, I don't think they reason that you were going to win and *should* be stopped.

I think the Facebook AI bot / approach is much stronger.

[A_Tin_Can]

Yes, that's right:

https://github.com/diplomacy/research/b ... per_v1.pdf

As a reward function, we use the average of (1) a local reward function (+1/-1) when a supply center is gained or lost (updated every phase and not just in Winter)), and (2) a terminal reward function (for a solo victory, the winner gets 34 points; for a draw the 34 points are divided proportionally to the number of supply centers).

So, the bots are unlikely to value stopping solos. It's also interesting that the scoring system they use is not available on webdiplomacy (and never has been). I didn't notice that last time I looked.

Reading this https://arxiv.org/pdf/2010.02923.pdf , It looks like the facebook effort is trained instead with an (also trained) estimate of the final SoS scores given the board position - (at least the one in that paper is, I think). I would expect that approach to lead to better behaviour at stopping solos.