No video
Diplodocus - The AI That Conquered Gunboat Diplomacy
In early 2022, Meta AI Research held a Gunboat Diplomacy Tournament in which human players would face off against various AIs. I was the highest scoring human in this tournament - but I didn't win.
In this video, I give my thoughts on the two AIs that beat me - Diplodocus-Low and Diplodocus-High - and go into detail on why they impressed me so much.
Paper here: arxiv.org/abs/2210.05492
Game links:
First Game: webdiplomacy.net/board.php?ga...
Second Game: webdiplomacy.net/board.php?ga...
00:00 Gunboat Diplomacy
03:30 The Meta Speedboat Tournament
06:45 The Five Keys to Gunboat Diplomacy
11:18 Diplodocus
15:34 Game 117
22:52 Game 100
31:02 Conclusion
Пікірлер: 38
Oh so many things to comment. So first of all, I miss the meta tournament. I played (I think) over a hundred games in 3 months and because of that, my skill level improved significantly. I'm also happy to be part of both games analyzed here. About the first game, I did not remember it, but I did remember you prioritizing Scandinavia over your homeland as England in some other games, which wasn't a strategy I was familiar with. I'm not sure if that has anything to do with this, but I remember you doing this in a game in which I was a particularly stubborn Germany, repeatedly and uselessly attacking your Denmark, and therefore inducing you to believe I was the bot and only encouraging you more to go for Scandinavia and leave Britain to France. About the second game, I did remember it. As the Turkish player, I think I was also worried Italy might be the bot and that's why I decided to prop you up. I'm not sure what would have happened if I allied wth Italy, but I had a feeling I would be better served by siding with you as my junior partner in an alliance. In the end, being 1 dot over Italy, I can't complain too much. Interestingly, I executed almost the same exact anti-Austrian stab as Italy in a recent gunboat game I played. Austria was, as far as I know, a very strong player, but no one expects Tunis to be convoyed in Austria and Trieste to move to Vienna after you build 2 fleets in 1902. I was not aware I was copying the meta AI, when I did it I thought it was a spontaneous idea, but the mind is complex and probably I had memorized something of game 100 in my subconscious. I was about to comment that I think the bot would do worse on boards constituted by 6 high level players (and I think some meta games were like that, with 6 players with lots of dipcoins countering the bot pretty well). But you did say (almost) the same thing at the end of the video. That said, it would do worse not because it would be outclassed, but because its skill level would be, more or less, matched by the other players. I think Diplodocus is currently the best gunboat player in existence, or at least in the top 5. Their results are astonishing.
Absolutely fascinating. Persuading someone is a difficult task in general, but even more so in Gunboat. The AI seems to have mastered wordless communication and pushing all the right buttons in humans. I guess its because thats all the AI has ever done. Kind of like putting on a blindfold and fighting against a blind person.
@solsystem1342
Жыл бұрын
I doubt it was trained against human players. Most likely these strategies are universal for intelligent agents playing diplomacy. The balance of different powers may sway depending on the meta but, things like the western triangle are a property of the game not human nature. Edit: they did use a model to try and imitate how humans play but, it was used only to reward the AI. With deviating from normal human play being penalized but sometimes still optimal for play against humans (which is what it is actually training for) (for example how it values Sweden and norway more than a human).
You can find the paper here: arxiv.org/abs/2210.05492 It's really hard to overstate how impressed I was with Diplodocus-High in particular. I'm not 100% happy with this video, because I don't think it conveys what happened in the tournament as well as a full series would have - but it became pretty obvious we were never going to have time to cover the whole thing, so it was this or nothing! Happy to answer questions about it in the comments if anyone has any.
diplodocus is a great name
@BourgoiDino52
Жыл бұрын
It’s a Dinosaur name!!!!☺️
@solsystem1342
3 ай бұрын
Rawr🦕
Re: the Denmark disband topic around 20:00 What if the bot was _also_ considering that their army in Brest was going to be force disband (or result in a build somehow), so it might have an opportunity to rebuild F DEN as an army? Yes, this idea is moot if Austria takes Munich. But I would still say it's the better play, because Munich retreating to Kiel is effectively the same outcome as this disbanding F DEN move?
@CaptainMeme
Жыл бұрын
I could definitely see that! If you do get the rebuild off it's a huge boon, an army is much more useful than a fleet in that spot. Bre getting forced is fairly likely, too. Does that possibility outweigh what Hel/Bal can do? I *think* only if you're considering how the disband will affect England's disposition, and you want England onside. If you're willing to leave open the possibility of going after England still, Hel and Bal both have the big advantage in enabling that - and even if you're not, Hel seems like it should be pretty effective on defence given that the French attack is focused around the lowlands right now. So - it's probably not just for signalling, but I'd still consider it a great example of signalling cooperation!
@pyxxy_
Жыл бұрын
@CaptainMeme pretty much agreed that it's also for signaling to England. Something something the best moves have multiple purposes. But I would also say that I doubt Germany would have a chance to attack English Scandinavia any time soon? Pulling even two units, the minimum to think about taking Denmark, would collapse their western line? So holding onto 3 fleets in a crabs-in-a-bucket fight with England to see who gets to die to France last...not promising. I find your point about preventing a French occupation of HEL to be more convincing, but even that it's a huge deal, I think? Because France losing the NTH is likely to retreat to YOR or ENG or something that protects their gains from England, rather than try to get lucky in HEL? HEL also wouldn't be occupied by a fall retreat there, nor a fall move there, because tactically it does nothing and would give England the chance to get back on the island, probably?
@pyxxy_
Жыл бұрын
Not that one should give up looking for your next center, which is likely DEN given losing BEL would be hard to retake, and MUN a pain as well, but I don't see it requiring 3 fleets. God I'm just writing paragraphs that repeatedly compliment the bot, huh. The move is at least 50% signaling for sure, you're 100% right about that. It's just also great tactically. Just wow. Can't say I would have found this move as Germany here.
@NoNameAtAll2
2 ай бұрын
its*
This just seems like magic
It's incredible how much shit those bots had to deal with and still win the tournament lmao, where's the bot's KZread channel ;)
Very interesting video.
@DiploStrats would you be able to link the games that you talk about in the description of the video?
@DiploStrats
Жыл бұрын
Done!
Intuitively I'd expect any self-learning algorithm to do better at imitation the more it plays with humans. Which is obviously tricky to arrange. Also, kzread.info/dash/bejne/X5qEtKd8ZqbPldY.html
@DiploStrats
Жыл бұрын
I actually asked about this during the tournament! The bot wasn't learning at that stage; it was in its 'final form' as it were, and was just being tested. It was trained on a human dataset though and 'High' meant it followed human strategy more closely, 'Low' meant it did what it thought was best without regard for that.
@TheGameKat
Жыл бұрын
@@DiploStrats Fascinating stuff. We have come a long way since the Paradox bots expressing their displeasure through "Urgh" and "Grrarrrg."
Hello, what version of the game are you using. Where can I find it?
@DiploStrats
Жыл бұрын
This is using the webDiplomacy Beta. You can find it at webdiplomacy.net, then click the Play Beta button at the top right once you're in a game. You can't play against Diplodocus currently, but there is an earlier gunboat AI (DipNet) which is available for playing against either by going to play.webdiplomacy.net/ or by starting a bot game on the main site.
@daniellop1
Жыл бұрын
@@DiploStrats Thanks a lot for the reply!
Great!
i want to see more games with bots like this
Magnificent, diplomacy players creating an AI takeover scenario just because they wanna play against good AIs
h-WHAT
Not sure about Diplodocus but not too impressed by your game style, sorry: abandoning your home centers is a no-no: you bound youself to have no build and make your fortress where the others want to expand. I just jumped as replacement for Austria into a (press, not gunboat) game where the Italian had done almost exactly that: take Tri, convoy the other army to Tunis and the first thing I knew is that Italy was not my friend. The stab Diplodocus made on you was almost exactly what I would have done in its position but, for starters I would never have grabbed Tri, because I think it sends the exact wrong signal: war.
@DiploStrats
Жыл бұрын
I'd always value an alliance over a home center, especially in gunboat where you don't really have other ways to communicate! It's definitely risky to give them - if you can get the alliance without it, that's much better - but if you don't open south with the Austrian fleet in high level gunboat, you are almost always dead in the water. Austria is a pretty brutal country to play. I'd agree that grabbing a home center generally signals war, but in the Tri case specifically it doesn't, at least at high level - because the Tri swap is pretty established in the metagame at this point. There are a few other cases (eg where you can support someone somewhere at the same time as you take it) but that's the big one in the earlygame. The other home center abandonment (the one as England) is far, far more dubious of a strategy and was probably a bad idea. But I give the reasoning for why I did it in the video!
@LuisAldamiz
Жыл бұрын
@@DiploStrats - I beg to disagree, very strongly. I've played high tier games and championships when it was mainly an email game and that's IMO a no-no. It's just iike the "Key Lepanto", which even Edi Birsan says it's a no-no (because only its inventor could be so extremely carebear as to keep it going till the end). As Austria, having an Italian army in Trieste, threatening your other two home centers is absolutely suicidal and there's no way about it, high level or not. You may want to pretend you're fine with it for the blink of a second, just because you have no other options but usually you do: give Greece to Turkey, kick Italy out of Trieste, call France and put Italy in such a situation that they have to play your game, not theirs.
@LuisAldamiz
Жыл бұрын
@@DiploStrats - In fact, allowing Italy in Trieste is giving them the opportunity to make the stab on Austria work, bc Italy's problem in the Russian alliance is that typically Russia gets everything (Trieste even) but IF Italy can grab Vienna first... then Italy is much more likely to exploit the Austrian collapse vs Russia and Turkey, become a Balcan lord in equal terms with Russia and Turkey (or even superiority) and probably even destroy Turkey right after Austria (with Russian and French help hopefully). Alliances are important but alliance is not the same as submission, it's the same as respect and clear boundaries.
@solsystem1342
Жыл бұрын
@@LuisAldamiz I mean, have you seen top level diplomacy now adays? Italy ends up holding trieste a long time for many successful IA's. It helps resolve so much conflict at the cost of one center.
@LuisAldamiz
Жыл бұрын
@@solsystem1342 - How many of those scenarios result in an Austrian victory, let alone survival? Austria is an extremely fragile power and allowing an Italian army not just to grab a home center but to directly threaten to occupy the other two is suicidal madness. Not even the Key Lepanto, which only has Italy having Trieste for a short time in theory, actually works. If anything it'd be much better that Austria occupies Venice with the fleet (not a threat to other Italian HCs), supports Italy into Greece (fleet) and then the happy pair proceed to a barely belated Lepanto, after Italy builds +2 on first turn (and so does Austria).