AI bots on webDiplomacy: Skynet edition!

Site announcements from the moderator team go here.
Post Reply
Message
Author
ppaquette
Posts: 3
Joined: Wed Nov 14, 2018 5:54 pm
Karma: 33
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#121 Post by ppaquette » Thu Sep 19, 2019 2:14 pm

Octavious wrote:
Thu Sep 19, 2019 9:52 am
Fair enough. My ally bot must have been in particularly trusting mood. I think it's worth empathising that I'm not a top level gunboater by any stretch of the imagination, but I guess I have a decent feel for the patterns of the game.

One potential weakness of the bots is that they didn't seem to recognise the danger of unit that had sneaked behind the front lines. In my game the French bot was doing a passable job of maintaining a stalemate line against me (Turkey) but allowed a fleet of mine to slip through a gap to the north Atlantic. This fleet was subsequently ignored, and sailed around happily picking up supply centres unchallenged until the game was won. It was a bit like lamping rabbits. The bots were utterly focused on the stalemate line and oblivious to everything else.

But yeah, for the most part they're pretty nifty
I would guess that units sneaking behind the front lines is fairly uncommon in human games. The bot probably doesn't have enough examples of that behaviour to properly evaluate the danger. It's a limitation of the training method, if you bring the bot outside of what it has seen, you never know how it will react.
15

Octavious
Posts: 3843
Joined: Fri Sep 29, 2017 4:16 pm
Location: The Five Valleys, Gloucestershire
Karma: 2605
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#122 Post by Octavious » Thu Sep 19, 2019 3:17 pm

ppaquette wrote:
Thu Sep 19, 2019 2:14 pm
I would guess that units sneaking behind the front lines is fairly uncommon in human games. The bot probably doesn't have enough examples of that behaviour to properly evaluate the danger. It's a limitation of the training method, if you bring the bot outside of what it has seen, you never know how it will react.
I confess to having a very limited understanding of AI, but that's an interesting thought and has a ring of plausibility about it. :-)
1

LeonWalras
Posts: 123
Joined: Sun Dec 31, 2017 8:13 am
Karma: 145
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#123 Post by LeonWalras » Thu Sep 19, 2019 9:27 pm

jmo1121109 wrote:
Thu Sep 19, 2019 3:24 am
LeonWalras wrote:
Wed Sep 18, 2019 11:29 pm
Out of curiosity, were PPSC games included in the training dataset?
Yes, PPSC was. Of the classic games played on webdip 21,987 were WTA, 32,245 are PPSC, 1,728 are Unranked, and 2,216 are SOS.
That's the largest number of games. I imagine bots trained without PPSC games might be more adept at stopping solos.

bo_sox48
Site Admin
Site Admin
Posts: 3901
Joined: Thu Jun 08, 2017 1:01 am
Karma: 2785
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#124 Post by bo_sox48 » Thu Sep 19, 2019 10:57 pm

Be mindful that the dataset is larger when you include games not played on webDiplomacy.
2

LeonWalras
Posts: 123
Joined: Sun Dec 31, 2017 8:13 am
Karma: 145
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#125 Post by LeonWalras » Thu Sep 19, 2019 11:30 pm

bo_sox48 wrote:
Thu Sep 19, 2019 10:57 pm
Be mindful that the dataset is larger when you include games not played on webDiplomacy.
I will not be drawn into a discussion on whether PPSC or PlayDip produces worse players :p Obviously diluting our PPSC games with games from other sites is good, but better yet to purge the games altogether if we desire bots that attempt to prevent solos.
2

jmo1121109
Lifetime Site Contributor
Posts: 1099
Joined: Fri Sep 29, 2017 4:20 pm
Karma: 2944
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#126 Post by jmo1121109 » Fri Sep 20, 2019 12:16 am

LeonWalras wrote:
Thu Sep 19, 2019 11:30 pm
bo_sox48 wrote:
Thu Sep 19, 2019 10:57 pm
Be mindful that the dataset is larger when you include games not played on webDiplomacy.
I will not be drawn into a discussion on whether PPSC or PlayDip produces worse players :p Obviously diluting our PPSC games with games from other sites is good, but better yet to purge the games altogether if we desire bots that attempt to prevent solos.
But, the majority of solo's on the site come from PPSC games to, so if you remove those games you're also removing the bot knowing how to win. It's give and take, and it's more important for the bot to learn individual scenarios, which more games does far better than just games of a specific type.
2

jmo1121109
Lifetime Site Contributor
Posts: 1099
Joined: Fri Sep 29, 2017 4:20 pm
Karma: 2944
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#127 Post by jmo1121109 » Fri Sep 20, 2019 12:17 am

Octavious wrote:
Thu Sep 19, 2019 3:17 pm
ppaquette wrote:
Thu Sep 19, 2019 2:14 pm
I would guess that units sneaking behind the front lines is fairly uncommon in human games. The bot probably doesn't have enough examples of that behaviour to properly evaluate the danger. It's a limitation of the training method, if you bring the bot outside of what it has seen, you never know how it will react.
I confess to having a very limited understanding of AI, but that's an interesting thought and has a ring of plausibility about it. :-)
I would hope since ppaquette is the person who made the bot and what he said is correct :-)
6

Octavious
Posts: 3843
Joined: Fri Sep 29, 2017 4:16 pm
Location: The Five Valleys, Gloucestershire
Karma: 2605
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#128 Post by Octavious » Fri Sep 20, 2019 9:49 am

jmo1121109 wrote:
Fri Sep 20, 2019 12:17 am
I would hope since ppaquette is the person who made the bot and what he said is correct
Ah, so ppaquette is webDip's very own Newton Crosby? Consider my hat duly doffed :-)
1

BobMcBob
Posts: 3150
Joined: Mon Feb 11, 2019 6:28 am
Location: Also Not Austalia
Karma: 2122
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#129 Post by BobMcBob » Fri Sep 20, 2019 10:10 am

Octavious wrote:
Fri Sep 20, 2019 9:49 am
jmo1121109 wrote:
Fri Sep 20, 2019 12:17 am
I would hope since ppaquette is the person who made the bot and what he said is correct
Ah, so ppaquette is webDip's very own Newton Crosby? Consider my hat duly doffed :-)
Yep, that's right. Except the bot hasn't escaped yet. Or at least, it hasn't taken over my account yet.

Also, congratulations to ppaquette for gettin the highest (as far as I know) karma to post ratio on webDip. 11 to 1 is pretty impressive. Even out glorious zultar does not have a ratio that high. We appreciate your work.

zultar
Lifetime Site Contributor
Posts: 90
Joined: Tue Mar 14, 2017 10:59 pm
Karma: 1012
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#130 Post by zultar » Fri Sep 20, 2019 8:07 pm

Another update! :-D

Hi all,

We are pleased to announce that Philip has retrained the 1v1 bots and now they are doing much better. They have substantially improved their playing ability. Although we should really pace ourselves and spread out the development so we can regularly get more +1s. ;)

Thanks, Philip. Can't tell you how much we appreciate this.
12

CptMike
Posts: 309
Joined: Sun Dec 31, 2017 8:12 am
Location: Liège, BE
Karma: 54
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#131 Post by CptMike » Sun Sep 22, 2019 7:38 am

I have made a second test.

I guess that Philip taught the bot there was a stalemate to hold above all [shaa]. I guess also that it was taught to try the "rush" strategy (?).

That was an interesting game by the way to learn about the way the bot plays.

Here are my comments regarding bot's moves:

* A04 in Med was a strong mistake and I fear this is due to the [shaa] and try to take Tun whater the short term consequences... That moves made the bot lose.
* A05 build phase: why damned destroying the fleed in Adr whereas there are 3 armies around StP ? [ The bot had lost but it didn't know so. This shows that the [shaa] is too strong.
* The bot is too predictable. He played 6 times the same orders around Mun and Ber. See S07 when I didn't support Mun with Bur, A07 when I left Mar unprotected.
* The exageration in the [shaa] is also marked by the fact the bot left me convoys 4 armies on his back just to keep and attack the stalemate.

CptMike vs IA Bot-2

Wusti
Posts: 399
Joined: Thu Oct 19, 2017 10:12 pm
Karma: 232
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#132 Post by Wusti » Sun Sep 22, 2019 8:22 am

jmo1121109 wrote:
Fri Sep 20, 2019 12:16 am
LeonWalras wrote:
Thu Sep 19, 2019 11:30 pm
bo_sox48 wrote:
Thu Sep 19, 2019 10:57 pm
Be mindful that the dataset is larger when you include games not played on webDiplomacy.
I will not be drawn into a discussion on whether PPSC or PlayDip produces worse players :p Obviously diluting our PPSC games with games from other sites is good, but better yet to purge the games altogether if we desire bots that attempt to prevent solos.
But, the majority of solo's on the site come from PPSC games to, so if you remove those games you're also removing the bot knowing how to win. It's give and take, and it's more important for the bot to learn individual scenarios, which more games does far better than just games of a specific type.
The second you start censoring any of the data, the more skewed the bots will behave - the largest quantity with all variants is the only option for the smartest bots.

CptMike
Posts: 309
Joined: Sun Dec 31, 2017 8:12 am
Location: Liège, BE
Karma: 54
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#133 Post by CptMike » Sun Sep 22, 2019 9:10 am

Another trial but in playing A instead of F. I wanted to test the capability of the bot to counter an attempt of Austrian to "rush" in '04. This "rush" is very risky when the adversary knows and realizes that you are trying it because this can only work if he makes some mistakes.

The bot didn't and I shifted to a A05 "rush".

The bot surprized me by its agressiveness, which is the right way to counter this opening but also in playing a different way than in former games. I would have expected from a human GoL > Tyr in A03, but not from the bot. The bot also guessed nicely Mao > Por in A04 but if he had played Eng > Mao, I was dead.

I was too confident.

CptMike vs IA Bot-3

tr1285
Posts: 59
Joined: Tue Oct 23, 2018 8:25 pm
Location: NJ, USA
Karma: 79
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#134 Post by tr1285 » Thu Sep 26, 2019 1:29 pm

zultar wrote:
Wed Sep 18, 2019 4:23 am
Update: please read!!!

Hi all,

After getting the feedback from many of you and further testing on our end, Philip and I talked and he has worked some more of his voodoo magic. There are now two different versions of the bots. One version is the original bot that you have been playing with for the last few days and another is a more aggressive bot that may take more risks and try other moves. In a classic gb game, half of the bots will be the original and half will be this more aggressive version (all random). Our hope is that you will see more variations in the mid to end games and where the bots will take some more chances to go for the win. :)

In 1v1 game, half of the time you'll get the original and half you'll get the new version.

Please try out several games and keep some notes of where you think you see awesome bot behaviors or some changes in how they play and post them here. It'd be best if you can provide links to the specific game(s), the phase/year where you saw some interesting moves, and why.
It seems to me this improvement has made a big difference. I'm regularly asking for embarrassing draws now (5+ way).

I'm not sure if it's because pure aggression is an advantage in the game of diplomacy or more because of the mix of aggressive and passive bots helps the aggressive bots make quicker gains on their neighbors.

zultar
Lifetime Site Contributor
Posts: 90
Joined: Tue Mar 14, 2017 10:59 pm
Karma: 1012
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#135 Post by zultar » Thu Sep 26, 2019 3:48 pm

I wanted to clarify that neither version of the bots is passive. The new version does take more chances and would vary its moves more, so in some ways, that may appear more aggressive, but the original version is not really passive. Although all this may just be semantic. :)
What does it matter when you get your ass kicked. :-D
2

CptMike
Posts: 309
Joined: Sun Dec 31, 2017 8:12 am
Location: Liège, BE
Karma: 54
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#136 Post by CptMike » Thu Sep 26, 2019 4:05 pm

zultar wrote:
Thu Sep 26, 2019 3:48 pm
I wanted to clarify that neither version of the bots is passive.
Yeah. I have seen Mittag "suprized" by a bot when he was in Mao and Bre was free... He played Mao > Bre and the f**** bot played Eng > Mao. Who taught that to the bots ?!

On the other way, next turn, the bot diggued its grave in Por where Mittag destroyed its unit...

yuuki
Posts: 552
Joined: Tue Jul 30, 2019 3:48 pm
Karma: 42
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#137 Post by yuuki » Fri Sep 27, 2019 11:51 am

After 27 games in my Gun-bot series - 26 played and 1 canceled (england in a game in which italy started by taking munich in 01, Russia opened north and France got three builds) - I have finally gotten a solo with each country.

9 games for England
7 games for Germany
4 games for Russia
3 games for France
2 games for Austria
1 game for Turkey
1 game for Italy

Turkey and Italy solos seem quite straightforward. Seems to me like Turkey should always try for the juggernaut when playing with bots. Italy should *always* ally with Austria when playing with bots. Very few exceptions here. Won both on first try in my series but have played a couple of other games and observations hold up quite well. Turkey may be the easiest country to solo with.

France needs to try to ally with England and be patient. It can be quite painful to solo with France, but it is not as tough as Germany and England. It's still a solid country to choose against bots.

Germany needs to ally with England to try and get the solo. I tried allying with Germany and Russia but usually it ended up poorly. Russia for some reason tends to go after Germany instead of England after taking control of Scandinavia, which human players rarely do since they usually think taking England first gives better solo chances and understand that Germany needs to keep France in check. England proved to be a good ally. Conceding Belgium to England was key as both France and England will fight you to the death when Germany gets Belgium early.

England was the toughest to solo with in terms of games played to get there. It is key to take Russia out of the north fast and to never let France enter the backdoor as bot France apparently loves going for Liverpool and for the wales convoy. Germany seems a better ally than France, though France is a good ally when Italy is attacking F.

Austria and Russia felt very much like in human games. Either you cruise to a solo quite easily or you get crushed very early on. Russia seems tougher to play than Austria which is also in line with typical games. If Germany doesn't let Russia into Sweden it's basically over.

The summary to me is that Germany and England are way tougher to play in bot games than in human games, and Italy and Turkey are easier. It seems like the moves bots make in northern areas are more varied and difficult to predict than in the south. I need to give Germany a couple more shots now that I found Germany's 'natural' ally in England for the bots. I think I struggled a ton because I hate allying with England in typical human gunboat games and try to avoid it as much as possible, so was trying French and Russia alliances too much. The France bots do not seem to understand Munich support move Paris to Burgundy as an alliance sign.

There also seems to be a reluctance by bot France to go after Italy when Italy is focused east on Turkey, leading to a diminished risk to try a Lepanto. There is also a weird obsession by bot Russia to keep the northern fleet instead of disbanding it even when losing the south and having little chance to get anything in the north.
2

Hazel-Rah
Posts: 2
Joined: Sat Sep 30, 2017 2:47 am
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#138 Post by Hazel-Rah » Mon Sep 30, 2019 3:21 am

gameID=249556

After helping Italy establish a good stalemate in the western Med, England unaccountably allows me into Munich and then Berlin in the final years when both should have been stopped. Definitely some PPSC stink on that one.

Still, I'm pretty impressed after all of two bot games, knowing they can only improve. Very cool addition to the site. Would be neat if in the future, since they're named, bots could get their own personalities - one is super aggressive, one is a carebear, etc.

Carl Tuckerson
Posts: 457
Joined: Mon Apr 08, 2019 8:23 pm
Karma: 316
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#139 Post by Carl Tuckerson » Sun Oct 20, 2019 10:34 pm

Is there any functionality yet for games with some bots and some humans? I have not figured out how to set this up if it exists :lol:
1

e.m.c^42
Posts: 6320
Joined: Thu Jan 11, 2018 7:00 pm
Location: Rated 0/5 Stars; ☆☆☆☆☆
Karma: 1726
Contact:

Re: AI bots on webDiplomacy: Skynet edition!

#140 Post by e.m.c^42 » Mon Oct 21, 2019 12:33 am

Carl Tuckerson wrote:
Sun Oct 20, 2019 10:34 pm
Is there any functionality yet for games with some bots and some humans? I have not figured out how to set this up if it exists :lol:
I'm in one right now, so it's definitely possible, but am not sure how to make one either lol

Seconding this question

Post Reply

Who is online

Users browsing this forum: No registered users and 3 guests