The Effects of Culture, Religion and Cores on AI Tendencies: a controlled experiment

[dharper]

After a discussion on whether culture affects the AI's tendency to target nations, I decided to test a few hypotheses in a controlled experiment. I invite you all to replicate my experiment by downloading my scenario. I am also interested in criticism of my technique; I believe I have controlled the scenario, but I may have slipped up somewhere.

Hypotheses​

(1) Having a core on a province owned by another nation increases the chance of an AI targeting that nation.

(2) Having the national culture of a province owned by another nation increases the chance of an AI targeting that nation.

(3) Having the same religion as a province held by another nation increases the chance of an AI targeting that nation.

Setup​

To test my hypotheses, I created a new scenario (called 'AI Tendency Test 1') in which only ten countries existed. One of these was the test subject (France), which was given many advantages to encourage it to declare war quickly. It was also set to 'default.ai' to avoid targeting any specific nation. Another was the observer nation (Zimbabwe), which was given to the player by default. The remaining eight nations were 'targets' for the test subject, representing all possible combinations of same/different religion, culture and cores (2x2x2 = 8). Each target nation had only one province, neighboured only France, and had the exact same taxvalue and population as the others. Each also had the state religion and state culture of its province. Reformed religions were allowed but all of France was Catholic (to avoid it spending time and money converting instead of waging war).

I used more-or-less historical nations for simplicity, but it should be understood that these countries do not represent their historical nations! All relations with France were set to -100 and France was given a casus belli with every target nation on day 1 by event.

In so doing, I hoped to eliminate the effects of alliances, dynastic marriages, sizeable armies and good relations on declarations of war.

24 games were played, each until France declared war for the first time, at which point the game was ended and that target given a point. 24 was chosen as a reasonably large number that gave each target an average of 3 times if declarations of war were completely random.

France (test subject): France owned many provinces, all of which were national provinces (had cores), were Catholic, and had French culture. All of them had tax collectors as well (in earlier tests France spent some time building them, which slowed down my experiment). France also had cores on the Huguenots, Brittany and Burgundy. The French AI was set to 'default.ai' and given a starting army of 50,000 as well as a starting navy of 20 ships (although they did not use them, France began building ships as soon as the scenario began, so I gave them some).

Burgundy (target): Burgundy owned Bourgoune province and was French Catholic, claimed by France. Culture & Religion + Core

Brittany (target): Brittany owned Moribhan province and was Gaelic Catholic, claimed by France. Religion + Core

Provence (target): Provence owned Provence province and was French Catholic. Culture & Religion

Navarra (target): Navarra owned Béarn province and was Basque Catholic. Religion only

Huguenots (target): The Huguenots owned Poitou province (to avoid them neighboring Brittany) and were French Reformed. Culture + Core

Lorraine (target): Lorraine owned Lorraine province and was French Reformed. Culture only

Flanders (target): Flanders owned Flanders province and was Dutch Reformed. Core only

Switzerland (target): Switzerland owned Bern province and was Swiss Reformed. Control - no similaries to France

Results​

France inevitably declared war in the first 9 months of the scenario. The number of times each target was chosen by France as its initial victim in 24 tests was:

Burgundy: 12 (50%)
Brittany: 10 (41.7%)
Provence: 0
Navarra: 0
Huguenots: 2 (8.3%)
Lorraine: 0
Flanders: 0
Switzerland: 0

Conclusions​

(1) In 100% of 24 tests, the AI chose to attack a country owning a core province. This suggests that using cores to help influence an AI's expansion is a useful strategy - assuming, of course, that the target country is in a weaker alliance (if any at all) than the AI and that no significantly weaker alliance exists!

(2) A country owning a non-state religion province was chosen less only about 8% of the time, suggesting that the AI gives these provinces a low priority.

(3) Provinces with the same culture were chosen slightly more often than provinces of a different culture (12 to 10); this suggests a slight tendency for the AI to choose same-culture provinces, all else being equal. However, the difference between these choices is small. More testing is needed to determine if this effect can be attributed to AI behavior or simply random chance.

Follow-Ups​

The effect of cores was so strong that the rest of this experiment may be flawed; it would be interesting to perform this experiment again without the core targets (Brittany, Burgundy and Huguenots) to determine AI behavior.

This experiment assumed all taxvalues to be the same; it would be interesting to increase the taxvalues of non-core provinces to see if this changes AI behavior.

It would be also be useful to perform this scenario again without Brittany and Burgundy in it to see how often the AI chose the Huguenots (core but non-state religion) over non-core provinces of the same religion.

 

[unmerged(31425)]

Very interesting expirement. And since France had CBs on everyone, it most likely means the difference is really the cores, and not merely having CBs. I suppose the culture and/or religion could have an effect, but it seems that the cores could be an adequate explanation enough for your results.

I suppose I can update my AI guide, pointing out your expirement, to show that AIs weigh core claims in DOW descisions.

 

[Rythin]

Hmm, Flanders was a French core and they didn't dow it because it was wrong-religioned and wrong-cultured? Hmmmmmmm, maybe AI isn't that stupid after all.

 

[Norrefeldt]

Very interesting test, we really need to know more about this!

To test the culture's effect all cores could be removed from Frances' neighbours, we would then have a test without the parameter that I regard proven by your test to have the strongest influence on the AI's decision.
The culture's influence ought to be discovered by that.

 

[unmerged(31425)]

Ya, can you run the same test isolating cores, cultures, and religions? Like only have some provinces have 1 of each of these in a different set of playtests.

 

[dharper]

Folks, I have something very odd to report.

You see, Rythin's comment was quite interesting:

Hmm, Flanders was a French core and they didn't dow it because it was wrong-religioned and wrong-cultured? Hmmmmmmm, maybe AI isn't that stupid after all.

I doublechecked and - lo and behold - I had left out the core on Flanders, invalidating my experiment. I apologize profusely! It was sloppy, even careless of me. But easily fixed. Or so I thought.

I immediately added the core, then tested the scenario again. And waited. Ten years passed, and France still didn't declare war, so I restarted. And waited. No declaration of war.

I checked their relations - all around -100, although some had gone up as high as -60 by the time I ended the game. I checked the casusbelli - all there. I checked the cores - they now had all the cores correctly. I checked the army - it was still huge and none of the targets had an army bigger than 2,000 troops.

I decided something had gone wrong with the random number generator and restarted EU2. Same thing - no war.

I doubled France's army and wrote a new AI for them that gave them 100 war. No declaration of war. 10 years passed consistently without them doing anything but putting down revolts, sending merchants and investing into land tech.

Finally, I figured I'd screwed up somewhere, so I retraced my steps, restoring my original scenario, without a core on Flanders. France declared war in May of the first year.

So I gave them a core on Flanders again. And they stubbornly refused to declare war for at least 10 years (I gave up).

This seems to be a consistent and reproducable bug (at least on my machine).

I can't figure out what's going on - is there some 'balance of power' operating here? Or does the AI have too many choices? Or is it simply some unusual coincidence?

I'm going to try limiting the targets to groups of 3 to see what happens, but I thought other people should be aware of this.

 

[unmerged(31425)]

Another thing you may want to do is make every enemy country's provinces have the same commodity and tax value, as their income may have en effect on the expirement.

 

[dharper]

Another thing you may want to do is make every enemy country's provinces have the same commodity and tax value, as their income may have en effect on the expirement.

Good idea; I hadn't thought of that.

I'm getting increasingly frustrated. For no reason I can point to, I am running into a brick wall with my experiments. Every one I make now - regardless of targets - has been unsuccessful. France simply doesn't declare war. It can see them; it has a large army; it has a casus belli on them; it has a core on some of them; it has bad relations (-100 to -200); it even has 100 war in its AI; none of these things is making it declare war! Instead, I sit in Zimbabwe and see France do things like built its navy to ridiculous proportions, declare itself Defender of the Faith, and spend money buying technology.

I've never seen an AI so reluctant to declare war! What is going on? What have I missed?

If I can't fix this soon, I'm going to try adding EVERY target to France's 'hit list' and see if that works, but I don't like the idea...

 

[unmerged(31425)]

Is the AI agression level different?

 

[Toio]

Question:

Who does france have in their AI file in the "AREA OF EXPANSION" , If its France only. then they most likely will not attack flanders as Flanders is part of the low countries bracket. If its not there then add low countries and let me know of the results.

I know this will annoy IDFL , but in my AI for POR , I have removed Iberia for POR and they have not attacked Granada or castile or Aragon in my last 15 games.
Tested 40 games, and initially POR attacked Granada 50% of the 1st 10 games, but as u keep restarting these games with the Iberia missing then the likelhood diminshes. So my overall games ie 40 has a total of only 10 attacks on Iberian neighbouurs.

Just a process you can use to determine AI thinking process.

 

[unmerged(31425)]

There's another thing I'm concerned about this test. Doesn't the countries with cores have their relations go down quicker? They may be having an effect too. Or does the CB, make them all go down at the samerate anyways?

 

[Norrefeldt]

There's another thing I'm concerned about this test. Doesn't the countries with cores have their relations go down quicker? They may be having an effect too. Or does the CB, make them all go down at the samerate anyways?

Would putting all relations at -200 help, or would they still increase for the non-cores neighbours?

I have no advise for you dharper on why France suddenly gets so pacifistic by a core on Flanders..

 

[DSMyers1]

Would it be possible that the engine has two equally viable options and cannot decide between them? I am sure there is some method for choosing...I think...

 

[unmerged(13277)]

What an interesting experiment. I’m hooked, please keep us updated.

As idontlikeforms mentioned you should equalize the provinces, so they are 100% identical. It does have a say in colonization, so it probably have a say in war too. Sea and landlocked province might also make a difference.
And location of the country, like BUG is inside France might also be a factor.

I think I will test along with you…

 

[unmerged(17856)]

My guess is that the explanation that the AI simply cannot choose between the core targets is the correct one.

To determine the effect of culture, remove all but three tags, keeping France of course. No cores on either of the two non-FRA tags (tags A and B). You'd have 4 conditions:

Controls
1) tag A - same culture, tag B - same culture
2) tag A - different culture, tag B - different culture
Treatments
3) tag A - same culture, tag B - different culture
4) tag A - different culture, tag B - same culture

Predictions: all other things being equal...
Ho: culture has no affect on AI DoW behaviour...
condition 1 = condition 2 = condition 3 = condition 4 (i.e. probability of DoWing a given tag would be 0.50)

Ha: culture affects AI DoW behaviour
one or more of the conditions does not equal the others

You'd need at least 20 replications, 5 within each condition.

Another thing to take into consideration is the possibility of hard-coded targets for tags. I remember there were a bunch of people who hypothesized this a long time ago in order to explain some very strange AI DoW behaviours that were largely unaffected by manipulations of various variables.

You could test for this by cycling through different tag pairs, A & B, then A & C, then A & D etc.

 

[dharper]

Is the AI agression level different?

I've tried with the default.ai and another I wrote to give them 100 aggression - neither has any effect.

 

[dharper]

Question:

Who does france have in their AI file in the "AREA OF EXPANSION" , If its France only. then they most likely will not attack flanders as Flanders is part of the low countries bracket. If its not there then add low countries and let me know of the results.

At the moment, nobody. I felt this would complicate the experiment as I wanted to see why some countries expand ahistorically - we were discussing ways to prevent Aragon from attacking northern Italy at the time.

Just a process you can use to determine AI thinking process.

Sounds useful - probably wouldn't help Aragon, but I could see places where it would come in useful.

 

[dharper]

And location of the country, like BUG is inside France might also be a factor.

If you think it's important I can move the targets so they are all on the 'outside' of France bordering the sea.

I think I will test along with you…

Please do! The more the merrier. I've updated my experiment to include similar goods (every target now produces cloth); let me know if you find any bugs.

Afterwards, I doublechecked the target provinces to make sure they all had the same income. But of course they didn't. Of course? The different state religions have different income penalties, so there's a slight difference between them.

 

[dharper]

By the way, an interesting, if off-topic observation:

While watching France, I noticed that although I gave them a starting army in Paris, they always did the same thing with it, without fail. They sent the infantry to Lyonnais (plains) and the cavalry to Savoie (mountains). At least, this is what I assume given the numbers (10,000/40,000). Every single time except when rebels appeared or it declared war. Interesting, but I'm not sure what it means.

 

[17blue17]

After you have tested and eliminated everything else then try making SPA and other nearby countries weaker. I have noticed that when I make SPA stronger then FRA just sits there as if waiting to be attacked.