AI bots on webDiplomacy: Skynet edition!

[Carl Tuckerson]

AI Austria is faithful to a fault to human Italian players that don't attack it right off the bat. I've run nearly the exact same strategy with Italy now, where I execute a slightly delayed Lepanto and show friendship to Jane while it gets entrenched with AI Russia and AI Turkey. Austria rewards this show of patience by never attacking Venice, even when I convoy both armies into Anatolia, and even helps me kill off Turkey sometimes by supporting me to Constantinople. Austria is then consistently caught off-guard by my building armies in Venice and Naples and loses 2-3 centers to my very telegraphed backstab every single time. I'm surprised Austria even wants to play with me anymore...

https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel
https://webdiplomacy.net/board.php?game ... #gamePanel

I went on to win six of these eight games, draw one at 17 centers when the AI demonstrated a stalemate line, and 4-way draw another that hadn't reached a clear conclusion because I was stuck in a holding pattern due to atrocious logistical positioning, but which I think probably ends in a 4-way draw or 3-way draw in the end anyway.

I think this specific behavior is slightly too exploitable. I don't know if this is the place we ought to report these things; if another place would be preferred I will resubmit there. But I'm highly confident I am not of the caliber of player to win 75% of gunboats, even if Italy is rapidly becoming my favorite country...

EDIT: Seems I've been preempted a bit and incorrectly referred to the users as the names of the bots, when the bots play each country at random. This is even more interesting to me, as it suggests this exploit is consistent across multiple of the bots (unless I happened to hit the same bot playing Austria eight times in a row).

[diplomat61]

If all bots are equal perhaps it is an indication of hard a particular country is to play.

The bots are distributed randomly, so no. They play all the countries.

Are the bots equal? If so then it would be the context in which they operate, i.e. the country they are playing, that would govern their success.

[Tugster]

I have played three times, once as England, France and Germany, solo every time. Bots show zero long term awareness of impending solos, continuing to attack their nearest neighbor even while I am passing through 14 or 15 SC on the way to solo. I am a very experienced gunboat player and the bots are way below the skill of an average human player. It is very easy to predict what they will do each turn.

[Maniac]

Can I be the first to thank the AI bots for carrying out this thankless task

[jmo1121109]

I have played three times, once as England, France and Germany, solo every time. Bots show zero long term awareness of impending solos, continuing to attack their nearest neighbor even while I am passing through 14 or 15 SC on the way to solo. I am a very experienced gunboat player and the bots are way below the skill of an average human player. It is very easy to predict what they will do each turn.

I see your success didn't continue :). Their play will vary game to game even as the same country.

[KalelChase]

Enjoying the Bots... Two observations.

This may scratch my 'pick-up game' gunboat itch, so I might play less Live games with real people.

I'm not sure I love the 'draw' once I'm eliminated. I'd like the bots to continue so I can see which bot wins.

Other than that this is a fun start. Thanks.

[jmo1121109]

Enjoying the Bots... Two observations.

This may scratch my 'pick-up game' gunboat itch, so I might play less Live games with real people.

I'm not sure I love the 'draw' once I'm eliminated. I'd like the bots to continue so I can see which bot wins.

Other than that this is a fun start. Thanks.

The bots do not understand stalemate lines, so when we were not forcing the draw some of the bots stalemated each other and made it to year 2600+ in game and crashed the server, which is partially why the site was so slow Friday into Saturday.

[Johnny Big Horse]

I played 2. One as Germany and got creamed by Russia. The bots were relentless.

The other game I was England, and killed both France and Germany by myself and at the same time. They never combined against me. But I still had to draw, as I was so focused on them, Italy and Austria took the East, and we stalemated.

You guys are really good to beat these bots. I always lose in Gunboat.

This is a great training tool. You can play the same country over and over, until you get it.

[tr1285]

I've never soloed a gunboat game before I played the bots. I played them and won as Russia. I would tend to agree with some others here they have some room for improvement. I found their moves to be repetitive and predictable and quite passive on defense. My German ally didn't seem to care that I was going to solo.

I'll try some more games though.

[bo_sox48]

Do the points the bots have indicate how good they are?

We should look at their GR next month to see how good they are.

Bot games should be excluded from GR if I remember correctly, and that should include the bot accounts as well.

[CptMike]

Once again I DO insist :
Thank you very much for introducing bot's on Webdip!

There will be a next step in future to improve the IA's skills ( I have read and understood that we can NOT improve this at short term ) but in 1v1 the bot really made numerous errors:

Trial : CptMike vs IA bot

[mhsmith0]

Once again I DO insist :
Thank you very much for introducing bot's on Webdip!

There will be a next step in future to improve the IA's skills ( I have read and understood that we can NOT improve this at short term ) but in 1v1 the bot really made numerous errors:

Trial : CptMike vs IA bot

yeah kinda agree w/ that. was def. a fun game but bot had some issues
http://webdiplomacy.net/board.php?gameI ... nocache=27
fall 1902 mun-bur isn't super obvious from austria but it's borderline gg for france if it hits. not really sure it did great in coin flip type situations, and it wasn't super hard to engineer a 1905 destroy on his berlin army just be continuing to mess around and create weird situations

[jmo1121109]

As previously stated, the bots were NOT trained for 1 vs 1 so their skill in those variants will be significantly less than in classic games.

[Squigs44]

Do the points the bots have indicate how good they are?

We should look at their GR next month to see how good they are.

Bot games should be excluded from GR if I remember correctly, and that should include the bot accounts as well.

Well, GR is still in construction (its development took a backseat to implementing the bots), so anything pertaining to GR gets a bit tricky.
The way GR is currently run, bots are treated as normal accounts. The way GR will be run eventually is that games with bots won't count. We may however use those games to approximate the GR of the bot accounts just for the fun stat.

[Tom Bombadil]

As previously stated, the bots were NOT trained for 1 vs 1 so their skill in those variants will be significantly less than in classic games.

I must be part bot...

[Carl Tuckerson]

@smart people: were the bots trained using a lot of games from the site generally or were they focused on specific users / games?

I have a hunch that the bots' behavior loosely represents the historical webDiplomacy metagame up until the end of their training. Their behavior might reflect general trends in the userbase behavior, which might have practical lessons for non-bot games.

I'm sorry there are so many caveats and hedges in that theory, it rides on an assumption I can't validate. Would be interested to hear the answer

[Squigs44]

@smart people: were the bots trained using a lot of games from the site generally or were they focused on specific users / games?

I have a hunch that the bots' behavior loosely represents the historical webDiplomacy metagame up until the end of their training. Their behavior might reflect general trends in the userbase behavior, which might have practical lessons for non-bot games.

I'm sorry there are so many caveats and hedges in that theory, it rides on an assumption I can't validate. Would be interested to hear the answer

So according to the research paper (linked in the site banner) they used over 150,000 games in their dataset. This includes games on webdip as well as some other sites. So no, the data isn't really targeted.

[Restitution]

What's the reason that the SL models were used here rather than the RL models? Were you not able to produce decent performance with the RL models?

[captainmeme]

What's the reason that the SL models were used here rather than the RL models? Were you not able to produce decent performance with the RL models?

This is answered at the end of the paper:

"Probably the most interesting result to emerge from our analysis is the difference between the SL agent (trained on human data) and the RL agent (trained with self-play). Our coalition analysis suggests that the supervised agent was able to learn to coordinate support orders while this behaviour appears to deteriorate during self-play training."

[LeonWalras]

Has anybody had a bot game in which Italy attacks Austria early? Just curious as the bots seem to love the Austria Italy alliance so far from what I have seen.

Given how popular the opening is on this site, I actually wasn't at all surprised to see Italy attack Austria right off the bat.

http://webdiplomacy.net/board.php?gameID=246397