The Big Question - How Does The AI Choose Which Units To Build?

[Ozymandias]

In reverse chronological order (makes it easier to scroll down and up before writing ) --

@BomberEscort -- your points re: the relative merit of "weaker" units with multiple HPs are well taken; however, in the one test I ran this way, one AI Civ built the 2HP unit 70% of the time whereas the other did so only 40% of the time (I'm discounting the Player Governor for the time being as, in almost all other tests, it built very differently than the "pure" AI -- although, to be complete, it chose to build the 2HP 70% of the time as well.

@Comrade (used in the non-Soviet sense )Rocoteh --

(1) Your B-52 experience sadly seems to reinforce the overall impression that the AI is truly clueless re: certain kinds of units ... Anybody want to bet that bombardment units, aircraft and ships are all "grafts" onto a vanilla-ground-unit-only AI engine?

(2) I'm delighted that you'll be re-running your test with 3 flags only per Civ Governor!

@ Der PH --

(1) Your experience with flags certainly serves to reinforce my own suspicion that the AI is built around the notion of a limited force pool at any given time.

(2) Your 6/4/1 experience certainly gives more credence that the AI really doesn't "understand" the value of higher MFs -- which again reinforces my suspicion cited above ...

(3) Good work re: the dive-bomber! One question though -- does it attack units or infrastructure ... ? And I wonder if the new Conquests flag of being able to attack a specific unit in a stack will cause similar build results and an attack-units-only approach.


--On a side-but-related note, in the thread on modding musket values, the few posts seem to indicate that the AI uses musketmen and pikemen together reasonably well -- i.e., generally uses one of each in a stack. This has not been my own experience, but I'm the first to acknowledge that mine might be anomalous. Anyone else? How about a general sense of the AI at least using O and D units wisely in concert?

Thanks Yet Once Again To All,

Oz

 

[Ozymandias]

Originally posted by BomberEscort


I wouldn't expect it to... the survivability of the unit has very little to do with movement. Movement gets your units to the battle faster and there is the retreat chance (insignifigant in comparison to other traits)... but other than those factors. Movement is practically useless, unless you are impatient for battle... I would think the AI would value a slow movement unit, 6/4/1 Grenadier to use a previous example, over the fast-moving cavalry. Since, in the long run, the grenadier will fare better than the cavalry...

In general, I agree, although I like the idea of at least a chance of retreat -- What about in conjunction with the "blitz" flag? Surely a higher MF would be valuable then ... ? (hmmm ... Note to Self: maybe try two equal value units, MF=2, give one the blitz flag ... and then repeat with - yes- MF=1 ...)

-Oz

 

[mrtn]

Oz, have you modded your ini file? If you don't add NoAIPatrol=1, the AI will often patrol it's lands. These patrols are often comprised of pairs, one offensive and one defensive unit.

BomberEscort: Since, in the long run, the grenadier will fare better than the cavalry...

I disagree. The cavalry have the option to retreat, so 66% of your (veteran) cavalry will retreat from a lost fight. This means that 66% of them will live to fight another day, you don't have to replace those casualties. This is far more important than you offensive units' defensive capabilities.

 

[BomberEscort]

Originally posted by ozymandias
--On a side-but-related note, in the thread on modding musket values, the few posts seem to indicate that the AI uses musketmen and pikemen together reasonably well -- i.e., generally uses one of each in a stack. This has not been my own experience, but I'm the first to acknowledge that mine might be anomalous. Anyone else? How about a general sense of the AI at least using O and D units wisely in concert?

I have modded similar changes and it does not seem to work for me... The AI seems to produce the 'best' attacker and 'defender' but unless there A/D values vary greatly (e.g.- Longbowmen and Pikemen) they will not pair. For example I have modded a skirmisher unit as the UU for the Americans and gave it 5.4.2 stats... The AI will produce them, but will not pair them with Riflemen... I hope certain pairings are not hardcoded, since I can never get modded units to escort or be escorted (the exception being modded land bombardment units)...

 

[Ozymandias]

Originally posted by mrtn
Oz, have you modded your ini file? If you don't add NoAIPatrol=1, the AI will often patrol it's lands. These patrols are often comprised of pairs, one offensive and one defensive unit.

-- I hadn't thought of that! Gracias!

-Oz

 

[BomberEscort]

Originally posted by mrtn
...I disagree. The cavalry have the option to retreat, so 66% of your (veteran) cavalry will retreat from a lost fight. This means that 66% of them will live to fight another day, you don't have to replace those casualties. This is far more important than you offensive units' defensive capabilities.

Assuming a Cavalry or Grenadier attacking a Rifleman on Grassland, both veterans...

The chances for the cavalry are as follows:

Win 36.23%
Lose 55.20%
Draw 8.57%

So, roughly 1 in 12 cavalry will retreat instead of die...

Now the grenadier...

Win 44.80%
Lose 55.20%

Also, don't forget that deep inside enemy territory those redlined cavalry will probably get whacked... The grenadier will win slightly more battles and will survive slightly more. The cavalry can retreat, but will probably get beaten in foregin lands (a regular Warrior has a 54% chance of defeating a redlined cavalry). The difference is not great, and mainly comes down to style of play in this case. If you are a blitz type player, the cavalry are better. I personally take my time with slow moving units, pillaging countryside, etc... Weakening units in cities with cannons, so I need a high Defense unit to protect those bombard units.

 

[BomberEscort]

Originally posted by ozymandias


...What about in conjunction with the "blitz" flag?...

The 'blitz' ability is only helpful when the difference between the 'blitz' unit attack value and the defenders defense is greater than two fold... Here comes the math again

Attacker and Defender are Elite...
(1) Tank 16.8.2 w/Blitz vs. (2) Riflemen, Town (No Walls), Plains, Fortified

The tank has roughly a 85% chance of defeating (1) Rifleman and 50% chance of defeating both (with the likely result being 1hp left for the tank), the Tank would subsequently be defeated easily after this battle... In all cases (2) Elite Tanks w/o 'Blitz' are better than (1) with 'Blitz' ability. The 'Blitz' really shines when you have a technological advantage, but doesn't do as well v.s. equals or technologically superior foes... In those cases give me (2) Elite Tanks w/o 'Blitz'... The 'Blitz' ability is one of those deceptive attributes (like movement)... On the surface it seems like it is very important, but actually does not help the survivability of the unit, which I believe the AI factors heavily into its decision, hence the non-building of artillery units (survivability is zero without escort)... BTW I have modded some land bombardment units to have an attack equal to their bombard (and increasing the Defense) and the AI builds plenty... Which tends to support 'survivability' as a factor in unit build choices. The only problem is it uses them for standard attack instead of bombard. My next experiment will be to increase the ROF to 5 or 6 and see if this has an effect (along with lethal bombard)

 

[Der PH]

Originally posted by ozymandias
(3) Good work re: the dive-bomber! One question though -- does it attack units or infrastructure ... ?

I don't know yet, since my current game is the first in which I tested this approach. Flight has only been discovered a few turns ago and I just figured out that the AI actually builds dive bombers.
I will begin a little war soon.
As soon as I've seen AI dive bombers in action, I'll tell you about the results.

 

[Ozymandias]

Originally posted by BomberEscort
The 'blitz' ability is only helpful when the difference between the 'blitz' unit attack value and the defenders defense is greater than two fold... Here comes the math again

My next experiment will be to increase the ROF to 5 or 6 and see if this has an effect (along with lethal bombard)

This could essentially be the definitive test to see if the AI truly "understands" bombardment power to more than a minor degree -- I'm looking forward to hearing the results!

-Oz

 

[Ozymandias]

Originally posted by Der PH

As soon as I've seen AI dive bombers in action, I'll tell you about the results.

Excellent!

Many Thanks,

Oz

 

[The Last Conformist]

Just checking in again to see if there's any new great discoveries. Well, I'll never understand that AI ...

Keep up the good work, people! I'll try and participitate in this thread again when I can.

 

[korn469]

oz,

I'm not sure how you are doing your testing but here is a suggestion, equalize all of your test units costs

try testing these unit's for example

4.3.1 40 offensive
6.1.1 40 offensive
2.1.5 40 offensive
1.6.1 40 defensive

and for your test AI's have one with it's only trait set to build lots of offensive units, one set to build lots of defensive units, one set to build both, and without either offensive or defensive selected

 

[Patroklos]

Hey all,

The chance of retreat for cavalry is not all that great, but think about this. How many times do you end up with a stack next to a city you failed to take, with the stack full of red lined units that won there last attack just barely. Which unit will get out alive, the slow moving Grenadier or the faster Cavalry.

Also, the beauty of fast moving units is not whether they will kill anything where they are going, but rather whether there is that much to kill when they get there. I routinely use high movement units to a) bypass the heavily defended areas and snag an undefended city b) circle around a city, sever the roads/railroads feeding defensive untis into it (planes are good for this) and thus keep the AI from stockpiling units in the target city when it figures out where that is.

I don't know if the AI appreciates it, but it should.

 

[Ozymandias]

Originally posted by korn469
oz,

I'm not sure how you are doing your testing but here is a suggestion, equalize all of your test units costs

try testing these unit's for example

4.3.1 40 offensive
6.1.1 40 offensive
2.1.5 40 offensive
1.6.1 40 defensive

and for your test AI's have one with it's only trait set to build lots of offensive units, one set to build lots of defensive units, one set to build both, and without either offensive or defensive selected

Hi korn,

All tests before the HP test were geared so that AF+DF = 10 with each unit costing 10 (basically I wanted to have each city churn out 1 unit / turn). Range of units was up to 7 AF or DF; with 2HP AF+DF=5. "10" obviously seemed a good number to base %s around ... Tests were done first with three units, all AF+DF=10, MF=1, HP=1, all flagged O/D; then 6/4/1 flagged O with a 5/5/1 flagged O/D and a 4/6/1 flagged D etc.

Best,

Oz

 

[mrtn]

Sorry to get off topic, friends.
BomberEscort: The grenadier and the cavalry both had 6 attack, so they should win as many battles. But, let's stop this, we won't change each others playing styles.

 

[Rocoteh]

Question to all who plays on Earth-maps:

Have you ever seen AI build great numbers of naval
units in the Mediterranean?
I added the masterpieces made by Colonel Kraken
and Smoking Mirror to my mod.
AI only built a few galleys, that was all.


Rocoteh

 

[BomberEscort]

I tested the modification to the artillery I mentioned in a previous post...

I had lethal land bombard on and gave them a ROF of 8... They still only bombarded with about 25% of their arty force. The rest they just parked in the nearest city and did not use. I modded the arty to 0(12).3.1 ROF=8. I fear the AI will never use bombard correctly. When I give them an attack of 12 they use the it for a standard attack and do not bombard.

I think my next mod will be to limit the range to 1. Maybe values greater than 1 for bombard range confuse the AI. I doubt it. But I will try this... More to follow...

I will test the following:
Artillery---
0(12).0.1 ROF=8 (25% Bombardment, 75% Parked in nearest city)
12(12).0.1 ROF=3
12(12).0.1 ROF=3 Range=1
0(12).0.1 ROF=3 Range=1

I will also test on open land and backing them 'into a corner'...

 

[BomberEscort]

Originally posted by mrtn
...BomberEscort: The grenadier and the cavalry both had 6 attack, so they should win as many battles. But, let's stop this, we won't change each others playing styles.

Last word, I promise The retreat percentage comes off of its win percentage since the determination to retreat is calculated between the 2hp-1hp boundary according to several 'retreat rules' ... Granted the overall retreat percentage is small. I still say those redlined cavalry will get whacked! :arrow:

That's it I'm making a mod Cavalry v. Grenadier... this will be settled once and for all

 

[Rocoteh]

"The rest they just parked in the nearest city and did
not use." BomberEscort

That is the problem with artillery in ACW:

AI will park it in the nearest city.


Rocoteh

 

[BomberEscort]

Originally posted by Rocoteh
...That is the problem with artillery in ACW:
AI will park it in the nearest city...

I have often thought that the AI has a limiting factor of some kind in the number of units it uses at one time. I have made modern invasions of his home continent with 32 Marines and they do not get effectively countered, even though he has an extensive RR network and 4 times as many units to make a counter attack. He seems to do 10-12 counter attacks and then stops for some reason. I feel this was put in to, keep the player from complaining that the 'AI whiped out my invasion force in one turn'. I remember reading somewhere on this forum that Mike from FIRAXIS or maybe it was Dan said that the AI could be made unbeatable and that the game was originally too hard... Maybe a mod could move this to General Discussions and make it sticky. Increasing the chance a Firaxian/Breakawayian /Atarian will see this post and comment on the 'innerds' of the AI choice of units and AI deployment.

Where are you when we need you Soren?!?