AI: Why is it so hard to improve?

[Txurce]

1UPT is definitely part of the problem, and IMHO other strategy games have done better with the AI, but let's just be fair. This is not an easy problem to solve.

It may not be easy, but it's been long solved by modders.

 

[GhostSalsa]

You'd think they could port most of the existing logic over, since Civ 5 and 6 have almost exactly the same combat model.

And then how would they patch it for new features? With coders who vaguely remember where they put this and this string 5 years ago? Better to have a new, broken system that is known to anyone

 

[Cromagnus]

It may not be easy, but it's been long solved by modders.

This may appear to be true, but there's not enough evidence to back it up IMHO. Hundreds of thousands of people have played the stock AI, including *all* of the best CIV players in the world, and some of the exploits and failings were only found due to that intense scrutiny.

Any given mod may get hundreds or perhaps even a few thousand players, but of those only a handful play it a significant amount. I'd be willing to bet that the total number of play hours of even the most popular AI mod is less than 1% of the total play hours of stock CIV. It's the same reason that Firefox was long touted as being more secure than IE. It was really only secure because hackers weren't targeting it as much.

So, while I admit I haven't played any AI mod enough to know either way, my suspicion is that it would fall apart under intense scrutiny as equally if not more flawed. Perhaps flawed in different ways, but IMHO you can't fix the core issues with 1UPT AI with a simple mod.

I don't mean to denigrate the amazing work people have put into mods of CIV. I've used many of them, and swear by them. I'm just saying that the evidence of AI Mod superiority is anecdotal.

 

[Deaga]

To be fair, I think the biggest "problem", if I can even call it that, is that having good AI simply isn't a priority for Civ. At all.

We here at CivFanatics are a very small part of the community. If you look around, most people will actually play on Prince/King, or even lower difficulties, and still struggle with the game, as it IS complex. For those people, just having the AI "be there" is pretty much enough. It doesn't have to be actually threatening, it just has to look or feel threatening. And I feel like it does that job well enough, given all of the "Strategy game of the year" reviews we've seen. Even if the AI has Warriors and Slingers by turn 300...

I honestly think this sucks, as the AI is pretty damn bad this time around, but it's understandable in some sense... Man-hours invested in making a better AI would mean less time invested in introducing a new Civ, new units, fixing game crashing bugs, etc, which are much bigger possible issues to majority of the player base.

 

[tekjunkie28]

I can't find a direct quote anymore, only text quotes, but the primary reason for the bad AI, Sid Meier thinks you all are losers, doesn't want to hurt your feelings and lets the AI let you win.



So basically it's a choice for him to see threads where players feel all superior and brag about how they thrashed the "dumb AI" and threads where players feel all butthurt about being beaten by a "cheating AI" give the game bad publicity and leave. And this thread proves Sid Meier is 100% correct in dumbing down the AI just so you could win.

YOU SIR ARE AMAZING. I remember watchign that speech and it made me sick. What an idiot to think people dont want a challenge....

 

[tekjunkie28]

Ill be honest the AI is this game is better then civ5 was on release. I like that its aggressive I just wish it would fight harder. I did accidentally loose my first game to the religious spamming AI and almost lost my 2nd but I declared war and mopped him off the map. He should have been a harder fight but it was a push over... THIS is the issue I have.

 

[dfrani]

Civ AI has been effectively torpedoed since the introduction of 1UPT. In Civ IV I remember being scared %^$# of Monarch level armies.

Yes! Civ I was hard! I felt accomplished that I could win on Prince. Part of that could be age, experience, and skill over that time period; but I remember it being very hard. Not so much lately...

 

[ThunderLizard2]

t may not be easy, but it's been long solved by modders.

Don't agree with this at all. I played a few AI mods in Civ V and although better they were still easy to beat and they were quite slow per turn - like 30-45 seconds - in the late game.

On a chessmaster scale, if an average player is 1600, the current AI is about 100 and the final Civ V BNW AI was maybe 900.

 

[TehJumpingJawa]

Moving goal posts will have a negative impact upon the AI too; if the mechanics of the game are constantly in flux, implementing & balancing AI logic to work with them will be impossible.

Looking at the mechanics in civ6, they have neither been thoroughly play-tested nor balanced, which points to the obvious likelihood of them being thrown together at the last minute.
Thus it's no wonder the AI doesn't possess the logic to effectively interact with the mechanics.

I really do wish they'd get rid of 1upt; not only for simplifying the problem for the AI, but for lightening the load on the player. (e.g. directing 10 units from one end of your empire to the other should NOT require re-issuing half a dozen move orders EVERY turn....)
From a UI perspective, even *Civ1's approach would be better than the tedium we have at the moment! (*infinite stacking, but if 1 unit loses on the defence, the entire stack is wiped)

 

[Unconquered Sun]

YOU SIR ARE AMAZING. I remember watchign that speech and it made me sick. What an idiot to think people dont want a challenge....

What Sid said actually doesn't matter at all. He was there when Civs I to IV were difficult, and he didn't care. He was right there in the year 2016 when Firaxis released XCOM 2 which is hard on the highest level and he didn't care. IMO the designers of Civ 5 and 6 had as much free reign as the designers of Xcom 1 and 2 when it came to difficulty levels.

I'd rate my Ironman highest setting Xcom 1 and 2 victories as harder to achieve than any of my CIV BtS deity victories, but easier than my CIV Warlords victories. I'd rate my Civ 5 and 6 deity victories as winning the second to easiest, if not the easiest XCOM setting.

Moving goal posts will have a negative impact upon the AI too; if the mechanics of the game are constantly in flux, implementing & balancing AI logic to work with them will be impossible.

Looking at the mechanics in civ6, they have neither been thoroughly play-tested nor balanced, which points to the obvious likelihood of them being thrown together at the last minute.
Thus it's no wonder the AI doesn't possess the logic to effectively interact with the mechanics.

I really do wish they'd get rid of 1upt; not only for simplifying the problem for the AI, but for lightening the load on the player. (e.g. directing 10 units from one end of your empire to the other should NOT require re-issuing half a dozen move orders EVERY turn....)
From a UI perspective, even *Civ1's approach would be better than the tedium we have at the moment! (*infinite stacking, but if 1 unit loses on the defence, the entire stack is wiped)

When it comes to return to stacks, I think you are in luck. It has already started, and not only with corps and armies, but with giving the AI a significant bonus to combat strength. All things being equal, deity bonus makes their units ~33% better on top of their purported 80% production bonus, 3 starting settlers, and 2 workers. Those are very significant early bonuses, in fact much higher than the ones in Civ BtS or Civ I for that matter, and arguably higher than those in Civ Warlords. Yet actual difficulty has tanked greatly which is attributed to 1UPT.

 

[Txurce]

This may appear to be true, but there's not enough evidence to back it up IMHO. Hundreds of thousands of people have played the stock AI, including *all* of the best CIV players in the world, and some of the exploits and failings were only found due to that intense scrutiny.

Any given mod may get hundreds or perhaps even a few thousand players, but of those only a handful play it a significant amount. I'd be willing to bet that the total number of play hours of even the most popular AI mod is less than 1% of the total play hours of stock CIV. It's the same reason that Firefox was long touted as being more secure than IE. It was really only secure because hackers weren't targeting it as much.

So, while I admit I haven't played any AI mod enough to know either way, my suspicion is that it would fall apart under intense scrutiny as equally if not more flawed. Perhaps flawed in different ways, but IMHO you can't fix the core issues with 1UPT AI with a simple mod.

I don't mean to denigrate the amazing work people have put into mods of CIV. I've used many of them, and swear by them. I'm just saying that the evidence of AI Mod superiority is anecdotal.

Let's move past your admittedly baseless opinion and on to your logic, which seems flawed to me. A huge sampling would be irrelevant, because most vanilla players are casual or, with all due respect, mediocre. (Just check how many play even King-level games.) How big a test do you need for a computer chess program? If it beats a handful of grandmasters, I have no doubt it's a serious program. For a mod that has been around for well over a year, being played by many vanilla Immortal and Deity players (including me), the fact that the AI in Vox Populi cannot be "gamed" is proven beyond a doubt. In fact, it gets tougher and tougher. Many VP players actually drop a level over time.

Don't agree with this at all. I played a few AI mods in Civ V and although better they were still easy to beat and they were quite slow per turn - like 30-45 seconds - in the late game.

Try the Vox Populi mod, then let me know how easily you beat it. As for late game speed... it's faster than vanilla.

 

[TruthfulCake]

I wonder where this idea that 1UPT AI is impossible to fix came from. Even without reaching Vox Populi (CBP) levels of competence, the combat AI right now only needs small tweaking to be threatening: they need to be programmed to attack cities when they can. Does that seem like an impossible code to people?

The fundamental problem with the combat AI right now is their inexplicable aversion to attacking cities. If you remove 1UPT and replace it with infinite stacks of doom with the same AI, you will still not see the AI as a threat because the stack will flat out refuse to attack your cities. Basically, the problem is not the mechanic - it's the AI itself.

The current AI got "massive armies" right. They got barbarian harassment right. Tweak the code to discourage pointless unit shuffling when in position to siege, and you get a credible threat. Invest many hours in refining the AI code like some group of unpaid volunteers have done for Civ 5, and you get Vox Populi AI.

This may appear to be true, but there's not enough evidence to back it up IMHO. Hundreds of thousands of people have played the stock AI, including *all* of the best CIV players in the world, and some of the exploits and failings were only found due to that intense scrutiny.

Any given mod may get hundreds or perhaps even a few thousand players, but of those only a handful play it a significant amount. I'd be willing to bet that the total number of play hours of even the most popular AI mod is less than 1% of the total play hours of stock CIV. It's the same reason that Firefox was long touted as being more secure than IE. It was really only secure because hackers weren't targeting it as much.

So, while I admit I haven't played any AI mod enough to know either way, my suspicion is that it would fall apart under intense scrutiny as equally if not more flawed. Perhaps flawed in different ways, but IMHO you can't fix the core issues with 1UPT AI with a simple mod.

I don't mean to denigrate the amazing work people have put into mods of CIV. I've used many of them, and swear by them. I'm just saying that the evidence of AI Mod superiority is anecdotal.

Civ 6 has been out for only two weeks. The CBP has been around much longer and even got a cool rename. If the hours people have invested into a game is the only measure of credibility in their assessment of AI performance, then I'd much sooner believe in the Vox Populi AI that many players swear by than I would believe in the early sentiments that Civ 6 AI is doomed.

Not saying that Civ 6 AI isn't bad; I'm just saying that if you so readily drew a quick conclusion about the Civ 6 AI in such a short time, maybe you should try out the Vox Populi mod for Civ 5 to get a quick taste of what a proper AI for 1UPT can offer. Then if you would still object to the 1UPT system, at least you would be able to present more reasoned arguments against it.

 

[GhostSalsa]

THANK YOU. Frankly, 1-upt needs to be removed from ALL games, this nonsense has gone on LONG enough. Chess. Candy land. Backgammon. Final Fantasy III. It's clear computer code will NEVER be able to understand having different things in different spots, and that this system is not popular with anyone but a tiny niche population…

1-upt has a self-evident appeal to players, it is the language of most of the enduring strategy games that have ever existed, it is worth it to include in a game like civ, and it is something computer code can do when the framework is fixed enough.

What V always needed to do and failed at doing, and what VI disappointingly still needs to do and has not tried to do at all, is identify the specific ways civ presents a non-fixed framework for basic logistical goals, and provide the AI with experimental solutions for that, until eventually the AI can move about the board with relative competence. I'all agree with any 1-upt hater that refining algorithms alone will never get anywhere in 3 years, it didn't with V. It takes trying new frameworks out, and VI hasn't tried anything out yet.

As the simplest examples, they made terrain and movement rules crazy complicated in VI. Just let the AI ignore these rules when not attacking. They made ranged too powerful again. Just make ranged work like anti-artillery and have no attack. Simple simple things, remove what keeps the AI from being able to get to combat.

 

[Nixalo]

The AI issue, from what I see, is no 1UPT.
1IPT allowed for fake difficulty.

The issue with the AI is that the AI does not understand a few key basic and key complex aspects of playing the game.

It's like a shooter where the AI bots don't use cover. So you give them double health. 1UPT is double health. You are letting the AI to change the gameplay themes of the game.

My guess is that most o the AI's problems are unintentional errors in their logic or bugs in some of way they get information. The AI isn't playing like a stupid person. It plays like a misinformed player. Like someone who are going off an outdated guide.

 

[Xmonger]

(professional software engineer whose worked in AI for decades (among another other areas) FWIW)

The first question would be what techniques would be useful here. Recent developments in deep learning come to mind. While the parameters explodes compared to say, a Go game (where a deep neural net with tree search system beat the best human player recently (AlphaGo)), I would be interested in seeing how well you could throw a DNN at civ. Having thousands of games to train on would be a necessary step. Except in small parameter spaces like city attacking I'd suspect that running inference might throw your average GPU under the bus. Certainly training would take some deep pockets.

One could consider a single layer NN doing some form of gradient descent with an appropriate goal function, as a good focused solver. It would probably behave predictably and stupidly much of the time, but might be entertaining.

Otherwise going back to classical AI techniques, a tree searcher with an optimization goal and hand coded knowledge db would probably be the default approach. Would work the best, until people figured out its boundary conditions.

So yeah, lots could be done here but as usual it comes down to time and money. I'll probably mess around with this in a mod for fun.

 

[Txurce]

(professional software engineer whose worked in AI for decades (among another other areas) FWIW)

The first question would be what techniques would be useful here. Recent developments in deep learning come to mind. While the parameters explodes compared to say, a Go game (where a deep neural net with tree search system beat the best human player recently (AlphaGo)), I would be interested in seeing how well you could throw a DNN at civ. Having thousands of games to train on would be a necessary step. Except in small parameter spaces like city attacking I'd suspect that running inference might throw your average GPU under the bus. Certainly training would take some deep pockets.

One could consider a single layer NN doing some form of gradient descent with an appropriate goal function, as a good focused solver. It would probably behave predictably and stupidly much of the time, but might be entertaining.

Otherwise going back to classical AI techniques, a tree searcher with an optimization goal and hand coded knowledge db would probably be the default approach. Would work the best, until people figured out its boundary conditions.

So yeah, lots could be done here but as usual it comes down to time and money. I'll probably mess around with this in a mod for fun.

Before you possibly reinvent the wheel, check out VP.

 

[TruthfulCake]

(professional software engineer whose worked in AI for decades (among another other areas) FWIW)

The first question would be what techniques would be useful here. Recent developments in deep learning come to mind. While the parameters explodes compared to say, a Go game (where a deep neural net with tree search system beat the best human player recently (AlphaGo)), I would be interested in seeing how well you could throw a DNN at civ. Having thousands of games to train on would be a necessary step. Except in small parameter spaces like city attacking I'd suspect that running inference might throw your average GPU under the bus. Certainly training would take some deep pockets.

One could consider a single layer NN doing some form of gradient descent with an appropriate goal function, as a good focused solver. It would probably behave predictably and stupidly much of the time, but might be entertaining.

Otherwise going back to classical AI techniques, a tree searcher with an optimization goal and hand coded knowledge db would probably be the default approach. Would work the best, until people figured out its boundary conditions.

So yeah, lots could be done here but as usual it comes down to time and money. I'll probably mess around with this in a mod for fun.

I have absolutely no idea what half of your post means, but I would like to know, in laymen terms, what that means in terms of improving the AI for a game like Civ. Since you seem like you are ready to make an AI mod for Civ 6.

 

[King_Course]

I can't find a direct quote anymore, only text quotes, but the primary reason for the bad AI, Sid Meier thinks you all are losers, doesn't want to hurt your feelings and lets the AI let you win.



So basically it's a choice for him to see threads where players feel all superior and brag about how they thrashed the "dumb AI" and threads where players feel all butthurt about being beaten by a "cheating AI" give the game bad publicity and leave. And this thread proves Sid Meier is 100% correct in dumbing down the AI just so you could win.

This closening remark, makes you exactly one of the people Sid is talking about. It's either/or, served in the most negative fashion.

 

[CIVLOYALIST]

Why is it so difficult to at least make them play 80% of us? It is the combat part that is the worst. They really, really suck.
Have Firaxis or who develops it now, admitted they don't have the skill to make it better or do they feel it improves? It is the main aspect I need changed.

I play CIV much less than you have had. But I've observed the difference between CIV 5 and CIV 6. Recently, I have played a couple of online games on KING and EMPEROR settings. The Barbarians perform exceptionally well, even on KING difficulty. They attack sensibly and "without mercy" as compared to CIV 5. One of the strangest moment in CIV 5 was that the AI barbarians could suddenly retreat from combat and it doesn't choose whom to attack if I placed 2 of my warriors around it. CIV 6 will eliminate the weakest and shows willingness to continue attacking.

Overally, I would say, AI is acceptable on KING and EMPEROR level.

 

[TruthfulCake]

I play CIV much less than you have had. But I've observed the difference between CIV 5 and CIV 6. Recently, I have played a couple of online games on KING and EMPEROR settings. The Barbarians perform exceptionally well, even on KING difficulty. They attack sensibly and "without mercy" as compared to CIV 5. One of the strangest moment in CIV 5 was that the AI barbarians could suddenly retreat from combat and it doesn't choose whom to attack if I placed 2 of my warriors around it. CIV 6 will eliminate the weakest and shows willingness to continue attacking.

Overally, I would say, AI is acceptable on KING and EMPEROR level.

The thing is, if the AI empires were anywhere remotely as aggressive as the barbarians, we would not have nearly as many complaints. They managed to make the barbarian menace real in the early game, proof that they can actually code an AI that poses a threat to human players. Why are the major civs' forces so comparatively shy?

Is it because the barbarians' sole purpose is to harass you? Is it because the AI for major civs have to juggle many other tasks? Do these even matter? Not sure why. But make the AI units all pillage-happy and not retreat-happy, and I guess we get a decently challenging combat AI.