France vs Austria Bot Challenge

[jmo1121109]

The game setup screen says "anytime you ready up your orders, the game will immediately move to the next phase" - but that is not happening in the game I created. I've readed my Spring 1901 orders and the Bot is just sitting there with a big fat next to it...

For reference, this was something I'd coded for the old bot. This bot is a newer one designed specifically for F/A and is performing better. It's currently under an improvement process though, so we're getting daily fixes from a specialized team. The tradeoff is some more wait times as they work through any issues, which seems well worth having a fully specialized 1 v 1 bot as the outcome.

[RoganJosh]

So if you play enough games, the bot will always win more than it loses. Much like how a rock-paper-scissors player that always randomly chooses an action with 1/3rd probability is unbeatable on average. We're clearly not there yet, but that's the goal.

Some nitpicking here: unbeatable is not the same a winning. To be unbeatable, you need a strategy without patterns for you opponent to exploit. To be winning, you need to exploit patterns in your opponent's strategy.

If you play RPS choosing actions uniformly at random, and your opponent knows this, then your opponent can just play Rock every time. If you have no mechanism in place to exploit the vulnerability in your opponent's strategy, then the expected outcome will be a draw.

That's the point of Nash equilibria, the the payoff is invariant of your opponent's choice. If the AI chooses orders by computing Nash equilibria, and then choosing at random, then you don't need to compute the Nash equilibria yourself. You only need to determine one option that has a positive coefficient in your mixed strategy. If the AI doesn't exploit patterns in your play, then you can just play that option. Expected outcome will be a draw. You can't beat it, but you're also not losing.

[Expected outcome will be a draw because of symmetry, so I'm assuming we're in a competitive setup were two parallel games are played to force symmetry.]

[Polynoamial]

So if you play enough games, the bot will always win more than it loses. Much like how a rock-paper-scissors player that always randomly chooses an action with 1/3rd probability is unbeatable on average. We're clearly not there yet, but that's the goal.

Some nitpicking here: unbeatable is not the same a winning. To be unbeatable, you need a strategy without patterns for you opponent to exploit. To be winning, you need to exploit patterns in your opponent's strategy.

If you play RPS choosing actions uniformly at random, and your opponent knows this, then your opponent can just play Rock every time. If you have no mechanism in place to exploit the vulnerability in your opponent's strategy, then the expected outcome will be a draw.

That's the point of Nash equilibria, the the payoff is invariant of your opponent's choice. If the AI chooses orders by computing Nash equilibria, and then choosing at random, then you don't need to compute the Nash equilibria yourself. You only need to determine one option that has a positive coefficient in your mixed strategy. If the AI doesn't exploit patterns in your play, then you can just play that option. Expected outcome will be a draw. You can't beat it, but you're also not losing.

[Expected outcome will be a draw because of symmetry, so I'm assuming we're in a competitive setup were two parallel games are played to force symmetry.]

Yes that's correct. I should have said "the bot will not lose more than it wins". Our goal is an unbeatable bot, but it might very well tie against some players. That said, in complicated games even figuring out what actions are best responses to a Nash equilibrium can be quite difficult (unlike in Rock Paper Scissors).

For context, I spent my PhD making AIs to play poker and approximating a Nash equilibrium is the route that led to superhuman AIs in that game. We didn't need to do any opponent exploitation. Even though the opponents knew our bot was non-adaptive, they were still unable to beat it.

[RoganJosh]

Yeah, completely agree. It's a question of whether the game is sufficiently complex for humans not to be able to determine best responses.

I think my opinion on FvA is that it's probably not sufficiently complex. I might of course be wrong (and I'd be happy to be proven wrong!). It's definitely a borderline case. There are some more chaotic 1v1 variants, such as Cold War, which are more complex.

[Thibauld]

Hi All !

I am not really familiar with the FvA game but i have it a couple tries when i first joined webdipl last year. This post made me want to try it again, here : http://webdiplomacy.net/board.php?gameI ... ocache=122

I must agree with the fact that the IA is not better than a human when there are no overly complex decisions or sets of orders to give. In the game i played, the control for Mun-Berl and for TyrSea-WestMed, to less an extent, was Key and gave me an easy win. I would think there were several better orders which would have put me in difficulty, for instance i would have expected a convoy to Livonia earlier and to put more pressure in the North.

I hope this helps and in any case it was fun for me to play !

Thibauld

[lyingbackstabber]

Well that was a hard fought victory:

https://webdiplomacy.net/board.php?game ... ocache=766

Took me nine tries...with Austria. After so many games, I started seeing tendencies that I was finally able to exploit. By that I mean, when it's down to a 50/50 guess as to which center to leave defenseless, it had tendencies that I started relying upon.

Nine games isn't a large sample size and there aren't a lot of options in 50/50 so maybe it's just me imagining it, but it felt like I knew which country they were going to go for. Ultimately, I think won this game bc of a couple of those lucky guesses.

But, I won...1 for 9 so far!

[Geofram]

Considering this was the first FvA game I ever finished, I'd say it isn't too bad: http://webdiplomacy.net/board.php?gameID=332417.

In my first game I made some clerical errors (working with one eye at the moment) and just conceded rather than play it out.

The bot has a really difficult time if you slip behind common stalemate lines, doesn't seem to understand convoy strategy either.

[goldfinger0303]

I have zero clues why that fleet in Brest didn't move to MAO to block Geofram in 1904. That seems like a monumental miscalculation by the bot.

Yeah, some stragglers behind stalemate lines can be dealt with and can be ignored (mostly if France sneaks behind Austria's lines)

But an Austrian fleet in MAO means almost a certain loss.

[orathaic]

Is the bot only analysing the board position? Or has it been trained on past human play and estimate it's best moves based on what humans in similar positions have done?

I think there was mention of learning from webdip... Also is it able to play other positions (like GvI) without extra training, or does it only know how to play specifics?

[Ezio]

http://webdiplomacy.net/board.php?gameI ... ocache=122
Thibauld's game.
A big question I have is Army Venice in Spring 1902. What is the advantage to supporting Vie-TRI? I could envision a situation where it is correct to prevent a Fleet Build in Trieste (if this were the Fall), but in the Spring, with no potential for a self bounce, this left me scratching my head.
I would assume better would be tapping TRI if you're throwing away the move, to prevent this very scenario.

[dipperjay]

A long and intense game, but I won in 1917. I had chances to win earlier, but each time the bot guessed right (or I played suboptimally...), until their luck ran out. It was fun.

https://webdiplomacy.net/board.php?gameID=332915

I miss the old bots, when I could play a casual game and win almost every time, just to feel good :)

[BobMcBob]

A long and intense game, but I won in 1917. I had chances to win earlier, but each time the bot guessed right (or I played suboptimally...), until their luck ran out. It was fun.

https://webdiplomacy.net/board.php?gameID=332915

I miss the old bots, when I could play a casual game and win almost every time, just to feel good :)

Just play GvI lol

[svkrIs]

Just won the bot 3 times as Austria with different openings, and would like to share some opinions.
http://webdiplomacy.net/board.php?gameID=332757
http://webdiplomacy.net/board.php?gameID=332996
http://webdiplomacy.net/board.php?gameID=333005

I'd say the bot is competitive, but not enough.
1. Does the bot only open with Mar-Pie, Par-Bur, Bre-MAO? I'm not sure if the AI can prove it's the best French opening, but as far as I know, you'll be countered in a FvA tournament if you always open the same, I did counter it with Anti-Alpine in game 2.

2. The bot seems not like to make a home-sc threat or defense, it never tried to enter Vienna/Trieste in all games, and didn't defend Burgundy much(and even lost Brest in Game 3), however, it seems to enjoy bounce Tyr/Sil/Gal very much. I'm not sure how you guys code it, but 'able to build' should be considered in value system.

3. The bot seems don't like to support hold its units, that's why I call it 'competitive' because a good French player should do this, I did learn something in its move, thanks to the programmers!

Will try as France in the future.

[President Eden]

Broke the bot lol.

https://webdiplomacy.net/board.php?gameID=333609

Game is still in doubt after I played myself and let Austria back into Tyrrhenian Sea.

[CptMike]

Broke the bot lol.

https://webdiplomacy.net/board.php?gameID=333609

Game is still in doubt after I played myself and let Austria back into Tyrrhenian Sea.

Don't forget Liv and Edi

[President Eden]

Broke the bot lol.

https://webdiplomacy.net/board.php?gameID=333609

Game is still in doubt after I played myself and let Austria back into Tyrrhenian Sea.

Don't forget Liv and Edi

I was trying to flex and win without taking them
I might have botched it though. Austria probably ends up with TYS and ION. Might have to quit showing off

[Mercy]

How might you have botched it? If you cover Tunis, support yourself to Bulgaria, tap Con with Ank and tap Smy with Eastern Med, your solo is guaranteed.

[President Eden]

How might you have botched it? If you cover Tunis, support yourself to Bulgaria, tap Con with Ank and tap Smy with Eastern Med, your solo is guaranteed.

I predicted my own failure to see this line
You are right of course.
Flex Status: Secured

[RayJay]

One should never loose that feeling that you play the robot:) As it was said earlier here, knowing what you do with Mun-Berl seems to be the key.


http://webdiplomacy.net/board.php?gameI ... #gamePanel

[Doug7878]

So, why has the bot stalled in this game ??? Mods ???

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