Basic question re AI: RTS versus TBS

[Growth]

• Feed all the players game turns into a massive online DB.
• Feed that to a Deep Learning Algo, the kind that are superhuman at Atari games, i.e. DeepMind.
• Improve on learning for a few months.
• Offer an online Subscription tailored to hardcore Civ players.
• Have the cloud/servers perform all the calculations and decision optimizations.
• Make Civ one of the most advanced Human vs AI experience on a PC ever since Chess.

When? Keep Dreaming!

 

[Pistol90]

With neural networks good enough probability for AI victory would be 99,999999999999999999% ... I dont think matematicly it is possible to be 100% , but in reality yes

 

[Aristos]

A categorical NO.

With that said, nobody here is asking for such an AI. I would be more than happy with an AI that can use ALL the game systems in an acceptable manner, and provide some challenge to my victory even with the bonuses included. We do not have that.

Anyways, we are far away from any AI that could beat a good player in Civ without bonuses and consistently (that is, without relying on the Binary Lottery of the RNG giving the human a horrible starting spot while some AI spawn in Valhalla).

With neural networks good enough probability for AI victory would be 99,999999999999999999% ... I dont think matematicly it is possible to be 100% , but in reality yes

Neural networks are not what you think they are.

 

[Pistol90]

Neural networks are not what you think they are.

I am not saying right now, but in 10 years period. Now its early-development, even prototype(experiment) in games -- but its most exciting and possibly important thing right now in whole science.

 

[Aristos]

I am not saying right now, but in 10 years period. Now its early-development, even prototype(experiment) in games -- but its most exciting and possibly important thing right now in whole science.

What I meant is that I don't think Civ is a "solvable" system. Chess is, although for the huge majority of humans (or all maybe) it behaves as if it isn't (and that's only 64 "tiles" and 6 different units for 2 sides)... Civ in comparison is a pseudo-open system, with too many variables and possible states (and increasing), that could be easily taken as a permanently open system, therefore unsolvable. If you think about it, it's designed to be unsolvable (that's why it is so addictive).

NNs excel in recognizing patterns; there is no pattern in a good civ game. In that sense, the devs of Civ 6 have succeeded in making it the most open of them all. NNs will not solve a system were there is no pattern.

The pseudo-pattern we have now, in civ 6, is the very bad AI. If they make it just a little more competitive (using all the systems available, for example), the challenge will increase tenfold (it will make the system more open, and weaken the existing pseudo-pattern)... translation: right now, if you do anything wrong, any amount of mistakes, you can count on easily defeating the AI in any war. That is the pseudo-pattern, and it detracts from an overall very good open design that should make the system unsolvable.

EDIT: solvable == an existing meta strategy that ALWAYS works.

 

[Red_warning]

I get the feeling that the random map generation of civ would screw with the AI a lot, in the case of SC II they have a preset map pool which is something that should help a slow adapting AI quite a bit.

 

[Krajzen]

I just wanted to say this is very interesting and informative thread.

In general, it is the great problem of all sophisticated strategy games - Total War campaign stage, Paradox Interactive grand strategy games, 4X, Civilization games with one unit per tile etc... Namely, that AI more or less sucks!

The best AI I've seen in big strategy game probably was Vox Populi mod for civ5, and it required God knows how many months/years of work by many people of whom at least one was actual proffesional programmer.

Don't get me wrong, I still think it is perfectly doable for civ6 AI to be significantly better if Firaxis hired more AI programmers and actually cared about the issue, but at some point I noticed that I cannot recally any big strategy game which AI was not at best tolerated by players.

The great tragedy of sophisticated strategy video games is simple: by their very nature they require very advanced AI for human players to have immersive and interesting obstacles, but they are also niche and not popular enough to not warrant massive innovations in this area.

If AAA action games, MOBA games, battle royale shooters etc needed as sophisticated AI as 4x games do, I think they would achieve more in this regard due to massive competition and massive resources their producers face.

 

[Pistol90]

What I meant is that I don't think Civ is a "solvable" system. Chess is, although for the huge majority of humans (or all maybe) it behaves as if it isn't (and that's only 64 "tiles" and 6 different units for 2 sides)... Civ in comparison is a pseudo-open system, with too many variables and possible states (and increasing), that could be easily taken as a permanently open system, therefore unsolvable. If you think about it, it's designed to be unsolvable (that's why it is so addictive).

NNs excel in recognizing patterns; there is no pattern in a good civ game. In that sense, the devs of Civ 6 have succeeded in making it the most open of them all. NNs will not solve a system were there is no pattern.

The pseudo-pattern we have now, in civ 6, is the very bad AI. If they make it just a little more competitive (using all the systems available, for example), the challenge will increase tenfold (it will make the system more open, and weaken the existing pseudo-pattern)... translation: right now, if you do anything wrong, any amount of mistakes, you can count on easily defeating the AI in any war. That is the pseudo-pattern, and it detracts from an overall very good open design that should make the system unsolvable.

EDIT: solvable == an existing meta strategy that ALWAYS works.

Well i disagree. I think Civ is solvable, maybe little more unpredictable than chess , but those unpredictions (if correctly calculated by AI from before) from human player can only harm human player - because AI would calculate best possible move in every curcimstances -- so you could only make same/or worse. You say you could count on easily defeat AI in war, but perfect AI would be able to fight a war like best Civ player in world --- that is probably most matematical part of Civ AI ... its just calculations on you attack/defense -> how much damage/healt there is done/left ... goal -> conquer as many cities as possible. At this would translate into cultural victory/science and so on ...

I think there is absolutely correct answer on turn 15 lets say should you build another warrior or builder dependent on how much info of map you have in that point (now if you say but if you could see whole map in that point than maybe answer would be different - yes - but that would be cheating then ...by AI or you) ... and if there is on turn 15 there is on turn 222

EDIT: just to say , there is percentege of luck in Civ (tribal villages lets say) so if that would be something which would (with you perfect play) bring you victory over AI that dosnt count

 

[uppi]

What I meant is that I don't think Civ is a "solvable" system. Chess is, although for the huge majority of humans (or all maybe) it behaves as if it isn't (and that's only 64 "tiles" and 6 different units for 2 sides)... Civ in comparison is a pseudo-open system, with too many variables and possible states (and increasing), that could be easily taken as a permanently open system, therefore unsolvable. If you think about it, it's designed to be unsolvable (that's why it is so addictive).

NNs excel in recognizing patterns; there is no pattern in a good civ game. In that sense, the devs of Civ 6 have succeeded in making it the most open of them all. NNs will not solve a system were there is no pattern.

The pseudo-pattern we have now, in civ 6, is the very bad AI. If they make it just a little more competitive (using all the systems available, for example), the challenge will increase tenfold (it will make the system more open, and weaken the existing pseudo-pattern)... translation: right now, if you do anything wrong, any amount of mistakes, you can count on easily defeating the AI in any war. That is the pseudo-pattern, and it detracts from an overall very good open design that should make the system unsolvable.

EDIT: solvable == an existing meta strategy that ALWAYS works.

You got it backwards: The reason why no AI, no matter how good, can ever achieve 100% win rate, is that there are games that are solvable. Suppose you play Sumeria on a Duel Pangea map against any civ without early game bonuses. On most maps, there will be a war-cart rush solution: A set of moves that a decently competent player (or AI) can come up with, which has no possible set of moves that can defend against it. No matter how good the opponent (human or machine) is, this will always result in a loss.

 

[Archon_Wing]

iF Georgia
THEN
Delete Settler and Resign

 

[Ziad]

Unfortunately great AI as expected by players is never going to happen by developers as the investment is simply never worth it. It's not really something you notice until you sink in hours upon hours into the game and move up to higher difficulties. At that point it's accepted that they provided you enough entertainment for your money.

Higher difficulties are just modifiers because those are easy to test and tune. They'll only really dedicate time to fix issues that arise during concurrent development. They only started caring about AI aggression towards CS when it interfered with Hungary for example.

Also AI is pretty universal. Better AI makes for harder games, and harder games make for frustrated players. They want the average player to have a good experience, and people who post on forums about civilization aren't exactly an average player. Have you seen the streamers? wew

AI is one example where it's probably left for the user to decide, and that's where mods come in. VP has taken 5 years to get to where it is today, and people still complain. Just gonna put that out there.

 

[Aristos]

You got it backwards: The reason why no AI, no matter how good, can ever achieve 100% win rate, is that there are games that are solvable. Suppose you play Sumeria on a Duel Pangea map against any civ without early game bonuses. On most maps, there will be a war-cart rush solution: A set of moves that a decently competent player (or AI) can come up with, which has no possible set of moves that can defend against it. No matter how good the opponent (human or machine) is, this will always result in a loss.

I disagree. In any case, you got it backwards. ANY solvable system will always be best/fastest solved by a proven algorithm, and as such, the more complex the system, the more a well programmed/powerful machine will be able to solve it. The human brain is not good at solving complex algorithms, but is excellent at abstraction (although it can be argued that the brain is, indeed, a pattern recognition machine). Your example is a little too limited, but even then, it does more to disprove your own point than to prove it. Any meta will always be solved faster by a well programmed, powerful machine.

As soon as we make the scenario closer to standard game conditions, the game can be considered not solved (in such a scenario, your example is but one of many steps to achieve victory). Other AIs could be doing the same, or could be using resources to develop instead of building a war machine, etcetcetc... by rushing your neighbor with war carts (under conditions of a better AI), you did not ensure your final victory, and may well have decreased your chances. Not solved at all.

 

[Krajzen]

Unfortunately great AI as expected by players is never going to happen by developers as the investment is simply never worth it. It's not really something you notice until you sink in hours upon hours into the game and move up to higher difficulties. At that point it's accepted that they provided you enough entertainment for your money.

They'll only really fix obvious issues or those that arise during concurrent development. They only started caring about AI aggression towards CS when it interfered with Hungary for example.

Also AI is pretty universal. Better AI makes for harder games, and harder games make for frustrated players. They want the average player to have a good experience, and people who post on forums about civilization aren't exactly an average player.

AI is one example where it's probably left for the user to decide, and that's where mods come in.

On one hand I agree. On another hand, Civ5/6 AI was in early versions so horrible that it was visible very quickly and a lot of people in internet reviews mentioned it... I mean, okay I don't expect civ AI to be "smart", but I expect civ AI to know how to play basic necessary systems and "how to capture a city" is very basic necessary system, or otherwise it removes the entire intended basic layer of "human player should be careful or he might be killed"

 

[Ziad]

On one hand I agree. On another hand, Civ5/6 AI was in early versions so horrible that it was visible very quickly and a lot of people in internet reviews mentioned it... I mean, okay I don't expect civ AI to be "smart", but I expect civ AI to know how to play basic necessary systems and "how to capture a city" is very basic necessary system, or otherwise it removes the entire intended basic layer of "human player should be careful or he might be invaded"

Oh I agree. There's a minimum to expect that's only fair, and Firaxis has dropped the ball on some counts.

*stares at ally destroying my city state*

 

[Pistol90]

You got it backwards: The reason why no AI, no matter how good, can ever achieve 100% win rate, is that there are games that are solvable. Suppose you play Sumeria on a Duel Pangea map against any civ without early game bonuses. On most maps, there will be a war-cart rush solution: A set of moves that a decently competent player (or AI) can come up with, which has no possible set of moves that can defend against it. No matter how good the opponent (human or machine) is, this will always result in a loss.

I dont know did you mean it in you post, but that makes another point which is it is impossible sometimes to win (I also include luck). At that point we cannot talk about AI that will ALWAYS win because it is impossible sometimes to win. But what is possible is to make best-possible solution --- and outcome will be win if this impossibility or luck does not intervene. In your Sumeria example best possible AI solution would be just make calculated number of units to defend (not possible if other player goes only for War-Chariot rush) and HOPE (because you played against human before and he sometimes dont use that strategy = so there is matematical possibiltiy) you can make them in time.

 

[Ziad]

I personally can't wait til such features are so easily implemented that developers would have to dumb down the AI for us mere mortals.

 

[Archon_Wing]

The problem with the Civ AI is not that it loses or it'll never beat a good player. Most people don't care about that.

The problem is the Civ AI simply doesn't really play the game. If this was a race, you'd be the only one racing and the AI would just be running all over the place trying to trip you over. On higher difficulties, they'd ride a car while you still run and they'd cross the finish line out of dumb luck still. Meanwhile, people brag about stealing the car and suddenly it's about that as a measure of skill instead of actually racing.

 

[Ziad]

The problem is the Civ AI simply doesn't really play the game. If this was a race, you'd be the only one racing and the AI would just be running all over the place trying to trip you over. On higher difficulties, they'd ride a car while you still run and they'd cross the finish line out of dumb luck still. Meanwhile, people brag about stealing the car and suddenly it's about that as a measure of skill instead of actually racing.

To be fair that's how even good AI mods like RTW2's DEI and Civ5's VP work. The AI is improved but they're still given a lot of training wheels. VP only just recently removed population and happiness bonuses AI receives.

 

[Archon_Wing]

To be fair that's how even good AI mods like RTW2's DEI and Civ5's VP work. The AI is improved but they're still given a lot of training wheels. VP only just recently removed population and happiness bonuses AI receives.

And there's nothing wrong with that.

Back when I did Starcraft AI scripts; we could only change what the AI built and when it attacked (even that was unreliable). We couldn't change unit behavior at all and thus even a bad player could win with the proper exploits. Of course, people that actually played the mod would just stay away from said exploits since there's little value in showing off how good you can exploit in a single player game.

For those that that wanted a better challenge though, the only that would work is Blizzard's "give money" command which basically gave the AI free resources on request. That is how Blizzard's insane AI worked. It does not play any better despite what they claimed, though it was design to take advantage of the extra resources.

So naturally many of us would follow Blizzard's lead and put those in a "challenge" version, but incidentally users inevitably would prefer the one that doesn't get any resource cheats. In the end, it just seemed that if the AI was playing a different game, it was a huge turn off.

Of course, even my non-cheating script cheated, and that was unavoidable. The AI cannot scout (go 1998 technology~), and thus I simply wrote the script to sometimes counter what the player was doing but threw it in a dice roll to make it less obnoxious and predictable. Sometimes it'd choose to make a counter; sometimes it didn't. Which was sorta needed because you could always do something on purpose to "trick" the poor bot.

Point is, it's all about how one maintains the pretense. We didn't have the tools at the time to really make anything

 

[Rath_O_God]

There are some notable differences between the AIs for GO, Chess and SC than what is needed for Civ.

1-in GO and Chess (and most of the games the super AI won against the pros in SC according to an article I read), the AI could see the entire field. They knew the position of every one of the human's units/pieces at all times. They know where the rook is, every line of attack that the queen can make, and thus can better manage its own moves so it doesn't open itself up to more risk than it needs to. This is not how Civ is ment to be played (even if you could get a spy in every enemy city chances are you will still have blind spots, and the entire first half of the game you have little to no knowledge what a random map will even look like)

2-SC is a different type of game than Civ. SC requires quick, repetitive decisions. Although the AI was limited on the APMs it was allowed to make, you don't have to worry about it slowing down, or hitting a wrong key. That is huge in any RTS game where you are not allowed to pause and examine things (something I do constantly in most RTS games I play because I prefer turn based game-play for the most part). Add to this the fact that SC maps are designed to be 100% even for both sides and you get an AI that can simply mirror what the top players have done in winning strategies and you get an opponent that can do everything that you do more efficiently.

3-The AI in Civ (as many others have pointed out) is intended more to be role players than an extremely competent opponent. The idea for single player Civ is you are creating and expanding an empire, and each civilization you meet is another interaction you have the opportunity to trade with, concur, or simply exterminate. You are writing your own history and the Devs want you to be Roosevelt meeting someone who acts (kinda) like Gandhi (only with more nukes). With that being said, the devs made the decision (for better or worse) to focus more time and resources on other aspects of the game and leave us with a (mostly) passable AI.

4-The AI has to actually run on the average computer. Some laptops already struggle with the game as is (mine for example, tried playing it at the airport the other week and boy was it chugging after a bit). Faraxis isn't going to invest in an AI that either a-takes 10 years to complete a turn on a mid-range laptop, or b-can only be run on a supper rig.

None of this is to say that the current AI could not be DRASTICALLY improved over current levels (even just an AI that isn't afraid to capture a city when it's been surrounded for 10 turns would be nice). And I'm sure if an upstart AI company wanted to start development on something that can really give the top players a run for their money they could, but I doubt it will make its way to a consumer grade release of Civ anytime soon.