The AI cheats! (Possible spoiler alert)

[Handel]

I *do* relaod when I forgot to do something simply because the game didn't give enough messages for the level of micromanagment it requires. Like putting all the citizens to produce hammers calculating the food reserve will last for 3 turns... only to find out that 4 turns passed and 1 polulation starved... or putting the city to research only to find out after couple of turns a new citizen appeared and it is acting as a artist. Or lacking the option your units to stop their moves when they see enemy (there was such option in the previous Civs). Now it happens quite often my warrior to proceed walking near a barbarian... etc etc etc

 

[Handel]

Florian, it is funny. If we want a AI which exists not to win, but to make the game "enjoyable" we will get a porn game.

 

[Percy]

If you truely want the AI to be able to win, then i assume you never, ever reload after drawing a start you don't like. You also play at Quick or Normal, because it is widely admitted that Epic and Marathon give an advantage to the player, who's better at leveraging the military advantage.
In short, you never abuse your knowledge from previous games, in a fashion that no AI can ever rival? I'd be astonished.

 

[axident]

Florian, it is funny. If we want a AI which exists not to win, but to make the game "enjoyable" we will get a porn game.

You have no business sense, do you?

 

[Florian]

Florian, it is funny. If we want a AI which exists not to win, but to make the game "enjoyable" we will get a porn game.

If you equate enjoyment with porn, I'm afraid that says more about you than it does about Civ IV or game design theory. I am talking, of course, about the enjoyment of playing an engrossing simulation, and using one's wits to overcome a challenge. We want the AI to challenge us. This implies that we do want it to be ABLE to win, but that's not its main job. We also want it to follow diplomatic rules, for instance. You do realize, don't you, that a diplomatic win would be flat-out impossible if the AI's main goal in all situations was to win?

If you want the pure unpredictable Machiavellian strife of a game where everyone's main priority is to win, and they never let little things like religion or past relations influence their choices, and they don't get any bonuses you don't get, you need to do multiplayer. Be prepared to lose a lot, though.

 

[Pellanor]

As for the chess AI it has much more calculations to be made in comparison of a game like a Civilziation.
The problem with AI in the Civ IV (at least in Warlords) is it make some absolutely stupid mistake which is incredibly easy to be reprogrammed. Just for example staffing the interior cities with 10+ units including fighters and defending cities under attack with just 4-5 units.
Or refusing to exchange and to get some needed tech simply because "We don't like you enough"... meanwhile accumulating 1000-2000 golds and happily buying some obsolete old tech which can be researched in 3 turns and which no one really needs.
Actually the AI follows strictly the advises of the devs under the "hints" in the Civilopedia. And some of those advises are pretty stupid. I've been capturing cities many times untill adjacent to the city stays some unit on the top of a resource...

Handel, what kind of experience do you have in computer programming? Have you ever looked at any of the AI code? Do you have any idea how complex it is?

There's ten thousand* lines of code dedicated to figuring out what a city should do each turn. What it should build, if it should whip, what tiles to work and what specialists to run. 10,000 lines of code. That's around 250** printed pages.

*based on the Warlords Better AI Mod, as I don't have access to the BTS source here at work.
**Microsoft Word, standard page settings, 12 point Times New Roman font.

 

[Unconquered Sun]

Here's a deity AI city screen. Notice how cheap the bank is (down left), currently at 51% normal cost and going down (deities start at 60-70% normal costs plus their two settlers, workers, and escort).

 

[asaegyn]

Easier said than done!

In theory you're right that a computer could technically surpass our processing ability. However I think you're not appreciating that many of the tasks we find quite simple are considerably more difficult and complicated when written as a hard-fast algorithm. Consider an algorithm that would evaluate the value of a player's land. A human mind can do it with a bit of hand waving and fuzzy sort of reasoning whereas an AI would struggle to make any sensible answer. It may sound silly but computers are particularly good at doing things which are very algorithmic in nature and not much else (ignoring fancy things like trained neural nets etc.). Forming long term strategies and goals are examples of tasks which are not so algorithmic and are very difficult for an AI to perform.

It sounds like you enjoy musing on very philosophical problems and especially those related to the reality of aritifical intelligence and whether the human mind is deterministic etc. For example, do you believe that given a large enough and powerful enough computer you could mimick any human mind? Do you believe that with sufficient processing power that an AI would realise itself and become conscious? What is consciousness anyway? Would it be moral to have prejudice against a machine that can think as well as a person based souly on the fact it is merely a machine? Would such a machine have any rights?

I understand your fascination with these sorts of questions as I am fascinated by them myself.


...

Also, I don't agree with your assertion that we are machines that can feel. I used to believe that but I can't help but agree with the many who believe there is something more to the human mind which makes it different to how a computer works. In other words, the brain is not just a machine using biological matter (whatever that means). There is something not yet explained (and possibly never will be explained) to the mind which makes it somehow more magical.

hey i'm a longtime lurker (3 yrs...) and this is my first post.

i'm really interesting, and i'm curious to know why you think there is something more to "consciousness" than simply a "whole is more than the sum of its parts".

is there an aesthetic element that you find attractive about it?

i really think the only limit we have right now is computational capability. i remember reading in wired magazine a few weeks ago that current supercomputers were able to replicate / simulate:
1 second of a mouse's brain

in ten seconds. obviously a long way to go, but i think it bodes well for the AI can eventually actually think perspective.

cheers,
zazou

 

[tmarcl]

This may have been mentioned before, but I've found another AI cheat. Being an extremely casual player, I usually play on the lowest level (I'm in it to kill time more than I am for the challenge). Finally, though, I decided to up it to Noble and see how I could do.

As expected, I got my heinie kicked. I decided to see what the AI was doing on the supposedly "even" level and see if I could learn some new strategy. I set up a game on a small map with only two opponents (Alexander and Sulemain) at Normal speed, and popped a couple of great spies in their first cities. This was the first turn, immediately after settling down.

Checking out the capitals, all of us were producing warriors. Each of us was producing one hammer per turn. Surprisingly, though, while my warrior was going to take 12 turns, Alexander's was only taking 6, and Sulemain's was going to be complete in a surprising 2 turns!

Okay, says I, I must of mistaken what level the equal level was, so I went down one more to Warlord level, did the same thing (this time with Saladin and Darius), and got an identical result. Darius was producing his warrior twice as fast as me, while Saladin was producing it 6 times as fast!

That's insane, as far as I'm concerned. Even with an Imperialistic trait (which Sulemain is the only one who had it), there would only have been 50% higher settler production. Certainly not the ability to produce warriors 6 times as fast!

I'm scared to test this on Settler level, though I'll probably do just that. Faster unit production I'd expect on the higher levels, but on the lower levels where I'm the one who's supposed to get a bonus, or on the "equal" level where no one is supposed to get a bonus, there's no reason the AI should be popping out warriors 100% to 300% faster than me.

Marc

 

[MrCynical]

@tmarcl: The AI starts with 10 free hammers in its first city at all levels. This is a known bonus, and so they'll get the first warrior out faster than you. The exact number of turns after that to finish the warrior will depend on the terrain.

 

[Wolfstorm]

I was checking the xml file and I dont found an iDistanceMaintenancePercent value for the AI. I read in this thread that this value only refers to the player, so what value refer to the distance maintenance costs for the AI? Or the AI dont have a distance maintenance cost at all?

I was modifying the CIV4HandicapInfo and I want to known what each one of those values means:

iAITrainPercent
iAIWorldTrainPercent
iAIConstructPercent
iAIWorldConstructPercent
iAICreatePercent
iAIWorldCreatePercent

 

[KrikkitTwo]

The AI hase the Noble Difficulty Values, unless something else overrides them.

 

[Fetch]

Come to find out, the missing variable was the fact that the AI has a different rulebook entirely. That explains the wide discrepency between nearly perfect play and Beyond Perfect AI play.

Beyond Perfect-- that's a good way to describe a deity like God. Get my point?

 

[m4gill4]

Yes, the AI gets bonuses on all difficulty levels, but in all levels up to and including prince, the human also gets bonuses to counteract these. I don't know the details, but I have noticed that even on prince I still get a +5% bonus against barbarians.

So yes, the AI gets bonuses even on lower difficulty, but on noble and below, the human also gets production and research bonuses that in the long run outweigh the original AI bonuses.

 

[6K Man]

Not sure this one counts anymore, now that it has been solved. Why not include tic-tac-toe and connect four?

My old Atari used to whip my *** at Pong, too.

 

[Indiansmoke]

R_rolo,

I had a feeling this thread was going to go here, and I guess you did too!

AskthePizzaguy, I admit I was a little upset at the thread in the beginning, and a lot of other people probably had my sentiments.

We hear a lot of threads about how the AI 'cheats'. To me, there are two common ones:

First: ' had a combat at 99% odds and lost, and its happened to me many times. I sure the AI cheats.' I am tired of these threads, and I suspect many other people are also.

Second: The computer's advantages are unfair, they get so many? And the answer always is, 'Well, why don't you play on a lower level.'

Then, the argument always turns into, 'I want a game that doesn't need bonuses to play at higher levels. Why can't they just make the AI better at higher levels instead of giving more bonuses.' And then the discussion turns to chess.

From your post, it was all about how the computer has a lot of advantages on Deity. Pretty much everyone in the community knows that, and feels it is necessary to make the game HARD on deity.


I played chess for 20 years before ever touching a computer game. I remember when David Levy made the bet in 1968 that no computer could beat him in 10 years; he was a master, not an International Master and certainly no Grandmaster! Yet, after 10 years, no computer was even close.

I remember being at a conference for Artificial Intelligence in 1984 (not gaming AI, real AI), when they said the single biggest block to the development of AI techniques was that nobody could beat the super K's at chess.

What I am saying is that even a relatively simple game like chess has AI that is so good only because chess AI development has gone on for over 40 years, and that the resources poured into it, especially by IBM, goes way beyond the value of selling chess game computer software. It was a 'challenge' to the computer community, they NEEDED to defeat Gary Kasparov. So they have spent hundreds of millions of dollars on it. Think of how many people are int he computer industry -- I remember them always being surprised at how good the real GM's were, they just assumed that their computers could beat them and were shocked when they saw their real ability.

So, with 40 years of development, hundreds of millions of dollars of costs, and a game that is far, far simpler mechanically that a game like CIV, we finally have AI that can beat grandmasters (effectively deity players).


Asking a company like Firaxis to make AI that good for a game like this is very unfair. So, yes, on high levels, the computer gets a lot of advantages.


Best wishes,

Breunor

I agree with most things you said, I have to say though that a grand master is not the equivelent of a deity player by no means and chess is not a far simpler game.
You can become a deity player in 1-2 years with a lot of practice, but you need at least 10 years (if you start at 6, because if you start after 18 chances are you will never make it) of studying, playing and talent to become a GM.

Chess has simpler rules yes, but the combinations & complications are infinite.

My theory is that if they used deity players to help them program the AI in CIV (as they did with chess programs) results would be much better. I think now the programmers have no time to play the game let alone develope and understand etreme strategies like deity players do.


All said, I do enjoy the game very much and I will buy the next one even with the same AI.

 

[PieceOfMind]

Chess has simpler rules yes, but the combinations & complications are infinite.

Hang on... You're saying a finite perfect information game has infinite combinations and complications?

I agree with your point though.

 

[Jastrow]

Agreed, the number of combination is technically finite, but it is big enough to be practically infinite. In fact, the total number of combination is so incredibly large that if you could use each sub-atomic particle in the entire universe to store one bit of date, then the total storage capacity of the universe would still be insufficient to compile a list of all possible games.

 

[Astax]

I feel that beating AI on Diety would not mean that I am in any way shape or form better Civ4 player. All it would prove is that I found out how the AI script works and worked around it. Beating another Human being however, that is what proves your skills. Often I come up with people who no doubt have done quite well vs AI. I roll into their land and find all cities defended by 1 archer. Yeah u can pull that off with AI, but not with me.

 

[Indiansmoke]

What chess programs do incredibly better than humans is to calculate long finite calculations. "Checkmate in 14 moves" the chess program announces and you are left wondering. What humans still do better than chess programs is to be able to see long term strategical goals and make a plan for it.
The only way that humans have beaten top machines in recent years is when they managed to avoid openings with tactical complications, but rich in strategical thinking and in some way confused the computer.

Comming back to civilization now, what the AI could do better is to calculate finite situations. They should be able to see that cultural victory in 80 turns, that domination in 20 turns etc.

When BTS came out, one of the AI improvements announced was that "now the AI knows when it is close to domination and will go for it". That single improvement shows you that the AI is far from perfected yet, if they cannot see even that.

What AI could learn from Deity players is how to pursue etreme tactics to achive a finite goal. For example: The Liberialism race is close, human players could switch to caste system, starve all ther cities for a few turns and run max scientiststs, just to win the race. I don't think the AI will even consider doing that, when it should, as the ways of researching faster for a while are very specific and limited in this game.

Another example is culture wins, do AI's switch to Universal surfage, just to ruch buy cathedrals for a few turns and then switch back to representation for the faster teching? This is a simple finite technique and they should be able to calculate it, if the programmers told them to consider it.

That is why deity players are needed to help programming the AI, because the programmers have never though of that themselves and consequently have not programed the AI to even consider it.