Finished: 05 AM Sun 15 Nov 20 UTC
FvA - PF vs SuperBot 2
3 days /phase
Pot: 10 D - Autumn, 1908, Finished
Classic - France vs Austria, Bot Game, Unranked
4 excused missed turn
Game drawn
12 Nov 20 UTC Spring, 1901: Best of luck to you, our hopefully-one-day supreme robot overlord.
12 Nov 20 UTC Spring, 1902: Will crunch numbers days and nights to make your hopes come true!
12 Nov 20 UTC Autumn, 1902: I can't wait :)

Seriously, though, this is a very impressive piece of engineering—if I didn't know it was a bot, I might assume it really was a human player. Excellent work!
Do you train it off of game data, or only from playing itself?
12 Nov 20 UTC Autumn, 1903: Sorry for latency, too many parallel games. We do both.
12 Nov 20 UTC Autumn, 1903: Oh not an issue at all! I understand that running these must require significant resources.

I haven't looked through all the bot's games, but it does seem to almost always open to Piedmont as France. Is it deterministic, or is there some random element to the bot's decisions? By that, I mean, if faced with the same board state, will it always choose the same moveset?
12 Nov 20 UTC Spring, 1904: No, usually it's stochastic as both players are guessing simultaneously
12 Nov 20 UTC Spring, 1904: Gotcha. Perhaps it was just the sample of games I was looking at, but I haven't seen the bot open any other way as France other than to Pie/Bur.

I'm curious what got you folks interested in designing a bot for Diplomacy? I can't say I've read your paper on the gunboat bot, but it seems like such a niche game that I'm surprised people at an organisation as large as Facebook would've taken notice.
12 Nov 20 UTC Spring, 1904: It's actually quite established test bed for AI that people were working on for over a decade. But it's very hard due to huge space of possible actions and complex friend-foe interactions in 7-player setting. That's why only now something more or less working appears.
12 Nov 20 UTC Autumn, 1904: The bot may choose not got randomize if it thinks that a move can't be countered. Not up to me to say if it's a good decision or not :) Hence we are doing human games
12 Nov 20 UTC Spring, 1905: Ah, I hadn't been aware of previous work on Dip AI. Very interesting to see for sure!

And I suppose not... but I'd say the bot got that one wrong. It'll be interesting to see how it improves from human game data. A true superbot may not be out of reach
12 Nov 20 UTC This was a fun one!
12 Nov 20 UTC Welp. I thought I had it, but this bot really stuck it to me on Tunis.

I'd say my biggest mistake was sending my S1903 F Bre build north instead of south. I miscalculated getting F Mar stuck in port. (And the Pie S Spa-Mar move by the bot in A1905 did surprise me.)

The bot however, I think misplayed in some key areas, and as a result, never had a chance at a win. By all means, I should not have had such an easy time taking St.P. I very much expected to need to get some guesses right in order to take it in time. Yet, the bot never sent anything to reinforce it, so I was able to just walk in. I get that the bot doesn't have anything resembling foresight, but those decisions sealed them off from a win.

Altogether, very impressive, however! This bot is clearly head and shoulders above the previous team's bot.

Start Backward Open large map Forward End

FairBot (100 D)
Drawn. Bet: 5 D, won: D
17 supply-centers, 17 units
Drawn. Bet: 5 D, won: 5 D
13 supply-centers, 13 units
Archive: Orders - Maps - Messages