This hit me time ago, but only now I write about it.

Deep Mind developed AI so sofisticated, that it beats best human players of Go single handedly. Now, I figure out Go is far more complex than Warlight, and introducing it (letting it learn how to play) would take this game to a whole different level. Level impossible.

We all get it that AI used in Warlight isn't brightest, so this might be interesting idea. Alpha go already mastered Go, played Apple Picker game and planned routes through London's Underground.

Since Alpha Go is BIG THING in the world right now, getting it over here would most definitely multiply number of users by... 100, 1000??

What do you think?

Warlight AI is open source, anyone can already develop a better AI for WarLight and submit it to be included in future versions of the game.

Now, I figure out Go is far more complex than Warlight, and introducing it (letting it learn how to play) would take this game to a whole different level. Level impossible.

The differents between Warlight and games like Go is that there is a luck factor. Eventhough a game might be 0% luck sr a game is far from predictable.
Go might scan all games played by 2100+ rated players and improve by playing against earlier versions of itself. However, as a human you can predict the AI will pick a ftb and counter it long term for example.
Further are on most boards no "Perfect" picks. A defeats B, B defeats C and C defeats A. Nevertheless Go would probably end up with a rating about 2100 but I can not imagine it would become a "Level impossible" force

Edited 5/29/2017 15:01:47

Well, that's where we disagree. Weather luck or no-luck, AI should be able to estimate position far better than any human. Go cannot be cracked through simple brute force, like chess was broken. AI has to learn that some positions are more favourable than others. Same would apply to WL.

All it'll take is to get attention of Deep Mind team :)

AI should be able to estimate position far better than any human.

This might be the truth for rather big maps with many possible picks but I disagree that Go would be able to evaluate a Strat ME board better than a human. There are limited options and a human, with enough experience, can come to the same, or even superior, picks. Also you should not forget about tactics. Often one correct prediction can win you the entire game. Sometimes pure luck.

Edited 5/29/2017 15:21:35

An advantage the AI might have, however, is that it is able to analyse a lot of your games. Due to a mass amount if data it will then have on your gameplay, it will most likely be much better at predicting your moves than any human might be.

That's cheating, both players need to have equal knowledge of eachother

Edited 5/29/2017 16:07:09

It's not cheating if both players can do it ¯\_(ツ)_/¯

You can go and analyse all games your opponent has played on the ladder and predict them on base of that. You can technically do it as well.

how am I supposed to analyze the games by an AI if it makes decisions based on his opponents past games

Edited 5/29/2017 17:59:59

LOL, this is going in wrong direction. Whatever you said applies to all games. You never played Go, obviously, so don't even try to comment on it's complexity.

Can you study your opponent's moves? Yes you can. It's just that unbearable fact that AI can do it much faster on unbelievable scale...

We're on wrong end of this bat...

No idea how many games an AI would have to play on ecactly the same map with the same distribution... but the distribution will change. It surely would take a while till the AI would be able to beat a good player on a regular basis

The prediction part of warlight is not really luck but gambling, which is a diffrence (you may still call yourself unlucky if you mispredicted, you couldnt have known i guess ;) ).
The gambling factor in a 2-player game like warlight could be nullified by finding the nash-equilibrium,
which basically means that there is a perfect strategy (meaning a strategy that wins at least 50% of games against any other strategy).
This strategy however wouldnt give exact instructions on what you do. It is more a distribution of chances to all not-dominated ways to play the current game.
This nash-equilibrium is impossible to be calculated by a human being and i doubt an AI could have enough power to do it on Strat MME for example (just an assumption).

But in theory, the perfect AI does exist. I am still not sure how well it would do.
Obviously, it is stronger than any other player out there in 1 to 1 comparison, as it would beat any player with a chance >=50%.
That doesnt mean it would do so well on ladders, as this AI wouldnt consider the factor of who its playing against (maybe you can discard a few non-dominated strategies out of the equation as you feel like your opponent isnt capable of finding those).
A Human may be capable of doing this and by that not just tie in the prediction game (like the AI does) but win it.

I dont know much about Alpha GO, but i think its an AI that gets stronger by collecting data from playing games, and not calculating much by itself.
Therefore i feel like it would have a hard time playing warlight, most likley learning very slowly.
Warlight has no open board and diffrent distributions every game, making it difficult to use the collected data.
There is a reason its known for GO and Chess i suppose.

Edited 5/29/2017 18:50:55

Hello

Deep Mind developed AI so sofisticated, that it beats best human players of Go single handedly. Now, I figure out Go is far more complex than Warlight, and introducing it (letting it learn how to play) would take this game to a whole different level. Level impossible.

I guess it depends on how you view complexity. The win condition of WarLight is pretty easy and is a mixture of primary income and then as some smaller factors position and armies on board. In WarLight there are no moves that backfire. There is no such thing as you hiting an enemy queen leads to you getting checkmate 3 turns from now. Breaking an opponent bonus is always a good thing, you just have to evaluate which of the dozen things you can do are more important and what your chances are at succeeding.

As for Alpha Go I don't really know how the bot works but I'm like 100% sure that the bot is minimaxing in the search tree and I'm also pretty sure that the bot is exceptionally strong at throwing away non promising branches (what probably is more difficult in Go than in chess). WarLight is more complex than Go in the sense that it has far more options each turn and with the players moving simultaneously minimaxing gets very difficult to give meaningful results. You can of course assume that you move first and then the opponent knows your moves and plays the perfect answer. However this will obviously result in your bot playing some nonsense play, far to passive and safe. In the WarLight AI competitions I only know of one player having claimed to have implemented some kind of minimax, however I don't know what he exactly did and his bot wasn't extremely great. That being said doing some local search on limited aspects of the game is certainly possible for a WarLight bot and mimicks the calculations a human does. For example when a human plans his expansion, he has some lookahead in his brain about what is possible in the first couple of turns.

Yes, the search tree for chess/go are absolutely nothing compared to warlight.

Map
Initial Position
EVERY SINGLE SETTING

I think that humanities AI knowledge could make a roughly 2300-2500 rated bot. It takes human reasoning at this point to play warlight at a high level.

A potential minimax tree for Go will likely dwarf any Warlight tree (as it does any chess tree, by a few orders of magnitude). Every turn in Go one can place a stone on basically any unoccupied space (19x19 = 361 total spaces), and even discarding obviously wrong moves for a large part of the game one will have around 200+ valid choices - each turn. Even if we consider attacking with 1,2,3,4,5,6...10 armies as different moves (as they are), a Warlight game on ME will not have so many different moves to pick from until very late in the game, and discounting obviously wrong moves (e.g. attacking large enemy stack with 1/2 of your large stack, or deploying all your income in one spot in backwater) there are usually even fewer moves to consider.

...however AlphaGo does not just do a minimax, as effectively doing that is impossible for Go (in addition to the size of the tree, evaluating a position is extremely difficult, unlike chess - and WarLight!). See any of the first few links @ https://www.google.ca/search?q=how+alphago+works. Unlike a decision tree or a min-max tree there is no easy human-readable explanation of what the values of parameters of the neural networks it trains mean. That is kind of the beauty of it, the people who programmed AlphaGo were not nearly as good a Go players as it is, and not just because they could not scan so many ply deep, but because they actually did not even know which strategies to use in which case which AlphaGo learned for itself. I'm sure if a similar algorithm is applied to WarLight (with some tweaks to the structure of the networks, and maybe some pre-/post- processing) it will do very well. Yes, it will take time to train it, but once it does (all games still being accessible in machine-readable form should simplify the technical difficulties of doing so) it should perform well

Edited 5/30/2017 20:55:40

Yes, it will take time to train it

that is the question and im curious

we should try this experiment with AlphaGo just to see how long it takes for it to become efficient at WL

if it ever does... that is

my bet is 2 years minimum whos with me

Edited 5/31/2017 14:49:48

@mslasm: It happens quite frequently that I hit an enemy stack with a smaller stack guessing it will move away in an earlier order.

What I want to say with this are two points:

WarLight has more (far more) possible moves than 361. However it's true that most of them can be seen as obvious nonsense and as for the move order also simplifications can be made.

The real issue is however that Go offers you perfect information while WarLights turns are simultaneous.

At theaigames there has been both WarLight and Go challenges. The bot winning the WarLight challenge didn't look a single turn into the future and I know of no meaningful minimax solution.

Edited 5/31/2017 21:48:20

I'd like to share a few opinions on this AI thing.

First, I'd say there are many different types of games on WL due to different settings, and the technique employed is very different from one to another.

Take a few for instance: local no-split, high luck percentage, 80/20 kill rates, 0 neutral, territories for income, multiattack. Not to mention coordination with partners and more. I think it could be a big trouble for AI to know what the best strategies are and decide what scenario it is undergoing.

Another big problem for AI would be the fog. AI may fare well in 1v1, but if it's playing in a 30P FFA it would be hard for AI to estimate who's stronger and who's weaker under fog, even if it's an AI employing machine learning rather than predetermined strategies and search, because AI cannot see, whereas humans can communicate in chats and private messaging.

Question is if an AI that is better then the best players is something good or something bad. Maybe for some competition where bots can battle each other it would be good. For in the game i have my doubts if it would good for the future of this game. I could just create games where AI does not surrender when one human remains and surrender on picks. I could turn the autopilot on in ladder games to achieve first place and probably more ways where people could abuse the advantage of a smart ai. If that becomes a structural problem the strategic community would get destroyed in a heartbeat