I'm sorry but....Combat results aren't fair!

[Iztvan]

Combat isn't fair! Last night my arabs attacked the Persians. I had cavalry & muskets, but he didn't have horses, iron or gunpowder tech...
And some weeks ago in another game, the b4$t@rd Xerxes sent 30+ immortals versus my poor & peacful Ottomans... I was crushed mercilessly!

The point is, combat isn't fair. Just make shure it's unfair to your opponent...

 

[Doc Tsiolkovski]

Sorry, I'm convinced the AI cheats occasionally in the game.
Of course, if you run the test with the RNG, it's absolutely fair.

But, the point is, since the RNG works like 3 turns for the future ('Preserve Random Seed' on), the AI Civs sometimes seem to know they'll get a favorable result.

Situation: The AI has a single Cavalry near one of your Cities with a Defender that is slightly better (with bonuses). Since the AI calculates its chances to win before attacking, it nearly never attacks here. But if it attacks, while you laugh about that stupid move....it wins.

But that is balanced by the fact that you warmonger may have tried it anyway, since you are willing to take the risk. That's a related topic to the AI Galley moves. The AI doesn't like to gamble, if it wouldn't sometimes know about the outcome, it would be further handicaped.

 

[Stapel]

Originally posted by Doc Tsiolkovski
Sorry, I'm convinced the AI cheats occasionally in the game.
Of course, if you run the test with the RNG, it's absolutely fair.

But, the point is, since the RNG works like 3 turns for the future ('Preserve Random Seed' on), the AI Civs sometimes seem to know they'll get a favorable result.

All has been tried a dozen times. If you really have your doubts,I suggest you come with some proof !

 

[Iztvan]

Ah, so you think the AI cheats by knowing in advance that a high-odds attack will succeed? Like taking a sneak peek at your opponents cards in poker is a different cheat from cheating when dealing the cards... a "pre-cognition cheat".

That would be impossible to prove either way. Shure, sometimes the AI succeeds in unlikely attacks... but sometimes the AI makes unlikely attacks and looses. And since the human virtualy never attacks when the odds aren't in his/her favour, we almost never see it the other way.

 

[Doc Tsiolkovski]

All has been tried a dozen times. If you really have your doubts,I suggest you come with some proof !

How would you try to test my theory/ impression?
As said, I know the RNG doesn't cheat in favor to the AI, and there is no effect on the big battles. But every few games, there's this one situation where you wonder why the AI attacks at all because the odds are really against it, since you were sure you secured that tile well enough, and -surprise, surprise!- it is victoriuos.
The point is not the outcome of a regular battle, it's the outcome of that single seemingly suicidal attack.
At some point you know what defense is needed to deterre the AI.
If you play a Mod with a wider variety of A/D values, like DyP, this is even more obvious. The AI will never attack a Rifleman (7/8/1) in the open with a lone Cuirasseur (6/3/3) under normal circumstances. But if it attacks (which is extremly unlikely, of course), it'll win.

 

[Stapel]

Well, I don't know how the AI decides its moves anyway....

I have seen the AI doing dumb things all the time. Sometimes they indeed get away with it, but most of the time they don't....

 

[Iztvan]

If the AI was able to predict the outcome of battles, it would never make any failed attacks, would it?

 

[Doc Tsiolkovski]

Sure, if the AI would know the outcome, it would never lose

My 'Theory' is more something like: What part of the programm tells the AI to attack?
Have you ever thought about how to programm an AI that isn't predictable without doing everything randomly? The Civ AI clearly has multiple personalities.

Some subroutines calculate the combat odds, and if it sees a chance, hurray. This is the case in regular battles, or sneak attacks, and the normal 'Field Marshall' of the AI. Would be the equivalent to the regular human player.

But if that would be all, the AI moves would be rather predictable, and boring. So I suppose something like a SpecOps Commander, who occasionally gets a hint from the RNG after the Field Marshall decided as expected. Only a small hint: "Chances are way better now, not impossible, like the Marshall believes!"
Equivalent: Galley suicidal runs. You know you have a chance, so you try. But the AI cannot gamble, period, forcing the programmers to mimick that trait somehow.

Also, there's of course the desperate last Lieutenant, who takes charge when things went really bad for the AI, responsible for all that suicidal attacking if you really outforce an AI Civ.
(Human player just before he quits "...,because the AI cheats!" )


That all isn't really 'cheating' in a way it has a big impact on the game outcome, it just makes Civ more spicy.

 

[Pembroke]

Originally posted by Doc Tsiolkovski
How would you try to test my theory/ impression?

Your theory is testable: If it were so, that the AI took "sneak-peaks" to the future, and _even_ if it did so only occasionally and randomly and not all the time it would _still_ show up in the long run statistics as the AI would win more often than it ought to.

Thus, a simple test scenario: make a mod where you have only 1 type of units available, say 200 units, and where all terrain is grassland and let the AI attack your similar units in any way it chooses and mark down how many times the AI wins and loses.

My prediction is that the AI will win just as many times as it ought to, i.e. there is no cheating.

Your theory goes also contrary to the "Theory Of Lazy Programmer." With that I mean why would an AI programmer _make_ the AI cheat in that way? It makes no sense from the point of view of a programmer. It doesn't make the AI decision logic easier in any way like "knowing" the map does (which is a proven cheat). In fact it makes it more complicated which is even more incomprehensible because you can achieve the same advantage (=more wins) by giving the AI a production bonus so it can throw more units against the player to compensate for its weak planning.

And the production bonus is there as a difficulty level setting.

Why do more work when you can achieve the same effect by doing less?

But if you still believe the AI cheats in battles the by all means run the simulations and show us unbelievers.

 

[Iztvan]

Doc T:
I totally understand what you are saying. Darned near impossible to proove either way...
There probably is some randomness in AI tactical desicions, if there is some knowledge of random seed involved... well well probably just have to take Firaxis words for it when they say there is no "combat cheat"... but then what you are saying isn't a "combat cheat" but a "pre-cognition cheat"...

Pembroke:
But if you make a mod with only warriors and grassland, all odds will be equal, and the AI wouldn't have any advantage if it knew beforehand the random seed (is that what you call RNG?), since it wouldn't be making a chioce as everything is equal.

But I think you are right: there is no AI combat cheat. The "lazy programmer"-theory is very convincing!

 

[Doc Tsiolkovski]

@ Pembroke: Sorry, but your model is absolutely not depicting my theory. I'm convinced it would show absolutely equal results.

What I'm talking about is not the outcome of a battle, but the chance of an unreasonable (in terms of A/D values) attack, that only seems to happen at all if 'someone' tells the AI the odds are quite good. Those combats a human player with his ability to reload would also try (if he uses save/reload). But the AI cannot use this strategy with ~90% loss propability in general, so there must be a small subroutine taking care of these rare occasions. And the lazy programmers...that is, in fact, a really convincing arguement. But how do you explain the well-know suicidal behavior of outgunned Civs?

 

[Pembroke]

Originally posted by Iztvan
But if you make a mod with only warriors and grassland, all odds will be equal, and the AI wouldn't have any advantage if it knew beforehand the random seed (is that what you call RNG?), since it wouldn't be making a chioce as everything is equal.

Actually I meant the following effect:

AI doesn't cheat and just sends 100 warriors one after the another to attack yours. Result: AI wins about 34 battles out of 100.

AI randomly (10% chance) peaks into the future to determine whether the battle will be succesful or not. If the attack will fail the warrior decides not to attack. Result: AI wins about 38 battles out of 100, e.g. 1/9 more than it should.

But I think you are right: there is no AI combat cheat. The "lazy programmer"-theory is very convincing!

Isn't it?

I'm convinced that when all the trappings and mumbo-jumbo is removed programmers are just a very lazy bunch of people who will spare no work to save work. Afterall that's what the job description is: make the machine do the job.

Besides, I should know as I'm one...

 

[sabo]

Has anyone asked Firaxis this question about combat results in one of their chats with us?

They were honest with us about the AI being able to see the whole map, I wonder what they would say to this question?

 

[Iztvan]

I'm with you. But, since the random numbers are generated in advance, and only get used up when something random happens, the same random number would be saved for the next attack desicion if the AI decided not to attack due to precognition.

So if the AI knew the random numbers in advance, it would stop dead in it's tracks and completely stop attacking when it knew that a bad number was coming up. This doesn't happen, so that's definately not the case.

Ok, so if the AI can "predict the future" in 10% as in your example? Well, it would skip attacking when it knew what was going to happen, but then the next warrior would probably attack instead.

Hmm, difficult to explain. And probably pointless to, since there most likely is no such AI cheat.

 

[Pembroke]

Originally posted by Iztvan
Ok, so if the AI can "predict the future" in 10% as in your example? Well, it would skip attacking when it knew what was going to happen, but then the next warrior would probably attack instead.

Ah, but skipping the attack is exactly why the AI would win more times in the long run.

The skipping warriors would not be a random sample of the attacks. They would be those warriors that would have lost for sure. If they skipped their attack then the outcome would be skewed in favor of the AI because of these "missing" lost battles. You just can't have it both ways: battle outcome as expected in the long run and the skipping cheat are impossible to have at the same time.

If the next warrior didn't peek and just attacked it wouldn't change this. Remember that random numbers are used for other things than battles, and when it's the AI's turn the random numbers are used for lots of things when it decides what it should do.

Also, just to include all possibilities: if the attack random numbers were separate from the rest so that if warrior #1 "predicted" it would lose and skipped the attack but the next warrior #2 attacked and lost for sure then you have to ask what was the point of having the cheat in the first place.

And now I notice that it isn't 1/9 more wins but less because the "peeking" warrior could also get the result that it will win the battle in which case it would attack. That won't change the fact that expected outcome of AI victories will be more than they should be. So this cheat could be found out if it existed. Which I doubt.

 

[Iztvan]

Random numbers used for ohter things than combat in the AI turn...
Yepp, didnt think about that...

Wellwell, i don't think there is such a cheat. Your argument just proves that it is possible to prove that it doesn't exist.

 

[Bamspeedy]

Well, actually, to back up what Doc Tsiolkovski was stating....

Barb activitiy was changed in PTW. Barbs won't attack if they have less than a 20%(?) chance of winning. Why do they sometimes attack a warrior on a mountain, and other times not? Theoretically, the chance of winning is the same at any point for a warrior vs. warrior on mountains. Why do they sometimes attack an injured warrior and other times not? AI units follows a different process sometimes, because an AI warrior may fortify and wait for reinforcements (especially if it was the last unit).

If the RNG spits out a number between 1-1000 (just as an example), and any number above 500 would mean a win, and any number above 200 would put the 'odds' above 20%. If the number was 250, then theoretically, the odds were above 20% so he would attack, but he still lost (because it wasn't over 500). If the RNG number was below 200, that's when he doesn't attack. Is this how the barbs decide whether to attack or not?
Examples:
RNG of 207-warrior attacks and loses
349-warrior attacks and loses
499-warrior attacks and loses
199-warrior runs away (but would have lost)
501-warrior attacks and wins
799-warrior attacks and wins
If the AI never attacks when the numbers come up as below 200, then that would skew the results. If it is supposed to be a 50/50 chance, but in 20% of the 'future' losses, they don't attack it ends up being:
20% run away
30% lose
50% win.
If you take out those run away cases, and just look at battles, then it would be
37.5% lose
62.5% win when it should have been 50% win/lose
And more importantly....Does the AI use this same reasoning/decision process?
If the AI is on defense, then there is no way they can cheat, and this has been proven.

 

[Doc Tsiolkovski]

Thanks for your assistance, Bamspeedy. You did find exactly that example I tried to get...not the 'Stack of Doom' against 'Pile of Defenders' scenario, but there are other, for the AI recognizable situations, where 'something' seems to help the AI to make the decision to attack or not.

Again, I don't call that a 'big' cheat, but it can explain why the AI sometimes attack, and sometimes not.

Also, I *feel* the random numbers are stored for a certain situation, not used one by one. Example (rather proofen, IMHO):

You attack a city. Now attack a second time, and the result will be the same, regardless how many units in the open you attack in between.
And, though this may be biased by memory - you only remember the situations were it was true - , if you attack a Barb and lose against the odds, then reload and skip the attack, you'll also lose in the defence during the AI's turn even if there's a better defender in the stack, like a Spearman.

 

[Veers]

Combat is fair and sometimes it isn't - - when the AI sends an Army and my MI defeats it, that's fair. But when a spearman beats my MA, that's not fair. Civ has always been like this and I think I'm used to it.

 

[billindenver]

Just yesterday I landed 8 veteran/elite Panzers against the Zulu who had only cavalry/infantry. I figured they might take a few HPs but the whole stack would probably survive.

Panzers should win 70% of those battles, but I forgot they'd keep taking hitpoints. The Zulus kept sending attackers in and wiped out the whole stack. I doubt they even lost 15 units.

Was it what I expected? No. Was it fair? Yeah, probably. It just pissed me off.