webdiplomacy.net

Awesome work!!

Temasek22 wrote: ↑

Sat Sep 14, 2019 4:34 am

I found out that the AI just stalemates itself eternally...

My game was Let's go bots!-2

Ah, you mean https://webdiplomacy.net/board.php?gameID=246292? That looks super funny.

peterwiggin wrote: ↑

Sat Sep 14, 2019 2:04 am

The drawback of this training method is that it requires lots of human examples. We can also train the models by having them play themselves and adjusting the parameters to make them more likely to capture centers and win. In this case, the two training methods yield bots with pretty similar strengths.

Would training the bots with each other not result in them developing a different meta than human players in variants with more than two players? I can imagine a bot that is trained to play against itself be very strong in a game with 6 bots and 1 human, but not in a game with 6 humans and 1 bot.

Wow. Well done to all involved.

@peterwiggin For a given power how is the sequence of orders determined given previous orders taken into account? Is it random, most adjacent neighbours, present or something else? Just curious.

Thanks

damo666 wrote: ↑

Sat Sep 14, 2019 10:47 am

Wow. Well done to all involved.

@peterwiggin For a given power how is the sequence of orders determined given previous orders taken into account? Is it random, most adjacent neighbours, present or something else? Just curious.

Thanks

It's a preset ordering that tries to put units that are close on the map close in the ordering.

This is cool, well done admins!

This very cool.

The bots do like a draw ....
As an example. 1 unit....
http://webdiplomacy.net/board.php?gameI ... ocache=726

Shouldn't that game have not been drawn based on the AI rules?

goldfinger0303 wrote: ↑

Sat Sep 14, 2019 5:12 pm

Shouldn't that game have not been drawn based on the AI rules?

Yes it should not have drawn. We obviously are having issues with certain cases where the bots aren't voting how we want them to. I'll reset their obedience levels to 100%...

Squigs44 wrote: ↑

Sat Sep 14, 2019 5:31 pm

goldfinger0303 wrote: ↑

Sat Sep 14, 2019 5:12 pm

Shouldn't that game have not been drawn based on the AI rules?

Yes it should not have drawn. We obviously are having issues with certain cases where the bots aren't voting how we want them to. I'll reset their obedience levels to 100%...

Currently it's only the top bot that's checked. In this game that was Austria, Austria didn't gain centers in 4 turns, so the game ended.

Just tried my first couple of 1-1s against the bots. Fun games. Thanks and congratulations, all involved.

I am imagining that the bots will learn (get better over time). Is that correct? Can you include them in the ghostratings so we can see their evolution over time?

The bots will not get better over time. They were trained once and do not learn from subsequent games.

jmo1121109 wrote: ↑

Sat Sep 14, 2019 7:34 pm

The bots will not get better over time. They were trained once and do not learn from subsequent games.

That’s what robot creators always say...

One thing I want to note is that in classic gunboat games, there is a fair distribution of openings the bots will use and I would play several games against the bots with multiple countries before coming to conclusion about how good or mediocre the bots are. As I have tested several dozen games, there have been games where I couldn't have done much else to beat back the bots, much less win, and there've been games where I won easily. For instance, if France opens against England and Germany doesn't bounce Russia in Sweden, and Turkey opens pro-Russia, it's damn hard just to stay alive as England. Same thing for Austria when Italy opening Tyrolia and Venice, Russia goes to Gal and Ukr, and Turkey pokes Serbia. In short, I would play multiple games as a single player just to see how varied the bots opening can be and how the game would play very differently.

On a separate note, I just wanted to say that it took quite a fair bit of work to get this update to roll out without it leaking. We didn't want to make promises that we couldn't keep, so kudos to everyone who kept it quiet and made it such a big surprise.

Bug report: Jane bit retreats all OTB when the first retreat is a forced disband.

Jane bOt. Gameid=246424 Phase = Spring 1915 Jane is Austria.Its A Bul could retreat toSer it Rum but it went OTB. See its first retreat was a forced disband in Mun

Sorry about the flooding. Now I note the same has happened to Kestaa Bot in the same Phase! NTH was disbamded whilst it had several options; while F Lon was forced OTB. Same game same phase lol

Please read my opening post, specifically this part:

"From a programming standpoint we cannot take suggestions on how to improve the bots play. You *will* run into situations where the bot's moves could have been better, you might have ideas on how the bot could be a better player, etc. We cannot accept any of those. The training of the bots is something done by the university partners and we do not have control over it. At a future point we might, and we do have some plans to help them improve the bots for 1 vs 1 variants, but I do not want anyone to get offended when we say we can't take your suggestions, it simply isn't possible."

Wow, this is awesome and totally stunning in terms of what this project must have taken! Congrats to everyone who made it happen!

Now do Mafia.

Since I was asked to clarify what type of bug reports can be accepted:

A bug is something where the bot fails to make a choice, or the game crashes. If the bot successfully makes a choice and the turn processes, then it's not something that can be changed currently. If the game is sitting broken then we want to know.

Congratulations to all persons who contributed to this bold endeavour. Naturally I find some of this AI boots caper incomprehensible yet exciting stuff nonetheless. The collaboration with a University is excellent to read of, this project l hope has given some of the University's ambitious young whippersnappers opportunities in this AI bots & algorithms for Diplomacy caper.
I'm confused with regard to AI bot players​ learning from experience in the games they will be playing against humans here at WebDiplomacy. I read a post by Jmo from which I infer that the AI bot players will not learn more , ie they are designed to have learnt to play Diplomacy, but won't learn/adapt more from playing games here at WebDiplomacy.
I had expected that the new AI bot players would learn/adapt as they played more games.
For example, if beloved Dipbro Jamiet99uk ambushed an AI bot player with his superb Silesian Surprise and then cruised to Victory would the Dalek, oops, AI bot player gain Silesian Surprise capabilities & then potentially use them in subsequent games it plays ?

Are we still allowed to allude to AI bot players with descriptive terminology such as Dalek(s) &/or Terminators in the Forum with more Decorum or is that considered offensive to their species ? .. Gosh, I hope I don't get more trouble for using those words.