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

[Rocoteh]

ozymandias,

I have now an early conformation from the
playtesting of my WWIII scenario Russia Demands Respect.

Before Conquests I NEVER managed AI to respond to this with
regard to Japan production:
Build often Off Land, Air and Naval. I have playtested
as India and attacked Japan (Yes India will be important
in RDR) Now Japan produces 1 Stealth Bomber and
2 Nuclear Submarines. I see this as the first confirmation
that AI really have improved with regard to production.

If you wonder Japan 3 cities??? Yes in RDR cities will
represent regions rather than cities per se.
This way I will hold down the number of cities, and I hope this
make the scenario faster and more user-friendly.
Forced Labour will not be allowed since China are
producing very unrealistic numbers of mech infantry
with forced labour.

BTW AI plays China extreme aggressive, and its very
hard to defend the eastern part of India.

Best Regards

Rocoteh

 

[Ozymandias]

Hi Rocoteh,

That's good news about the build flags

BTW I also concur with your approach re: "regions" for cities -- to try to get a 1:1 real world:Civ population distribution and density is not achievable.

Best Regards,

Oz

 

[The Last Conformist]

[dance] Yea! Good news on the AI build choices front! [dance]

I'm starting to like more and more, and I'v not even got it yet!

ObOT: It doesn't really belong here, but as I mentioned I was going to test it, I thought I say I did nail down the "non-advancing Leader creation bug". It turns out that if the battle-created unit has a DF>0, a victorious attacking elite will not advance into a tile vacated by the defeated defender, but the battle-created unit will appear in it. This is certain for PTW; I'll of course re-run the test when I get C3C, but I rather doubt it's been fixed.

 

[Rocoteh]

Originally posted by ozymandias
Hi Rocoteh,

That's good news about the build flags

BTW I also concur with your approach re: "regions" for cities -- to try to get a 1:1 real world:Civ population distribution and density is not achievable.

Best Regards,

Oz

ozymandias,

Yes, that is what I mean: If you really regard the "cities"
as real-area cities they will be absurd on most maps.

Best Regards

Rocoteh

 

[Rocoteh]

ozymandias,

GOOD NEWS!

From playtesting my WWIII scenario I have those stats:
India (human) in war with China (computer).
China have Off Land, Def Land , Naval and Air flagged.
China building after 5 turns:

8 Modern Armour
7 Mech Infantry
7 Stealth Bombers
1 Worker
1 Wonder

There is no production of naval units, on the other hand
Indias fleet is no threat to China....

I have started to place US units now. After that I will playtest
as US. Chinas fleet will probably be destroyed in 1-2 turns.
It will be interesting to see if AI responds with building
naval units.

Japan with Off Land, Naval and Air continues to build
Naval units only. As I mentioned before Japan has only
3 Cities/Regions. Before Conquests I was never able to
force AI to build Naval units in this way for Japan.

Best Regards

Rocoteh

 

[Ozymandias]

@Rocoteh -- Sounding very promising indeed!

-Oz

 

[Rocoteh]

ozymandias,

Have you given up this thread?
If so:

Sad!

The thread is still needed.

Best Regards

Rocoteh

 

[Bobby Lee]

Ok, Admittedly I havent read the whole entire post here....been gone a long time and im relatively tired tonight so I dont feel like it. That means what Im about to may possibly be flat wrong based on results u got sum time after page 4 or may have already been looked at and thought to be the way it works but here it goes none the less....

Is it possible that the offensive and defensive flags simply tell the computer what stat to look at to determine a unit's value. For example, If you flag a unit defensive then the computer looks at its defensive value verses other defensively flagged unit's values and decides based on that, then if it is a tie it moves on to its attack value to break the tie and then so on to other values. In the event it isnt a tie then it only looks at the attack values as an afterthought at most? The movement points would weigh in there sumhow...perhaps as a bonus (each movement counts as 2 point to add to its stat or sumthin funny like that?)

Perhaps that whole idea is just too logical to be possible for the Civ AI but...shoot, was worth a try, hope it helps sum!

Yet again GL and GJ!

 

[Rocoteh]

ozymandias,

I made a question concerning this thread.

You did not answer.

I suppose:
No answer = You have given up the thread.

Once again: Sad.

Anyway I wish You a Good New Year.

Best Regards

Rocoteh

 

[Ozymandias]

Happy Holidays Rocoteh!

-- No, I've not given up on this thread; I've just been hit with the triple shock of (1) heavy work demands (2) the holidays and (3) my girlfriend moving in on a "temporary" basis while her apartment is remodelled ...

I am behind on this thread; my own mod (actually 2, but I'll stick to one for now); and updating the URL library (I have a backlog of some 30~40 there).

-- Here's hoping for a calmer 2004!

All Best Wishes,

Oz

 

[Ozymandias]

OK, Gang, I'm reviving this thread & effort.

Personally, I'm curious to see how the AI weights differing flags against one another. I'm thinking of utilizing the old standard unit (5/5/1 HP=1) and modifying it so that there's one variant with each of the "significant" flags (ZOC, Hidden Nationality, etc.) checked and seeing what the beast does.

Any thoughts / recommendations / observations / requests / volunteers?

All The Best,

Oz

 

[Rocoteh]

ozymandias,

Very positive that you have activated his thread again.

The last weeks I have worked with scenario-design and
research, but I will report as soon I observe something
of interest.

The thread is really needed. Firaxis continues to run the
old Avalon Hill "mystique" policy on these issues.

Best Regards

Rocoteh

 

[Wladislaw]

I have noticed a couple of things about the ai unit selection, that surprised me a little bit.

When I first introduced the special forces unit into my mod it had a high attack value, bombard ability, blitz, all terrain as road, in other words it was a super unit, and the cost was not in line with the abilities. The ai built these like crazy and almost beat me with them.

In another mod, I forgot to put a cost on the composite bowman, the ai built them but did not fully utilize them. Of course it would take a human to create a bunch of one pop cities with zero production for the sole purpose of cranking out a zero cost composite bowman every turn and annihilating those ridiculous immortals. Persia had the capability of building 4.1.1 composite bowman (with C3C defensive bombard) for 0 but instead focused on 1.4.1 lancers for 30, and 3.2.2 ancient cavalry for 40. This tells me they don't really calculate value/cost, although maybe the formula can't handle 0 cost (divide by 0??). Might be interesting to give something a cost of 1 and see if the ai utilizes it.

 

[Ozymandias]

Originally posted by Wladislaw
I have noticed a couple of things about the ai unit selection, that surprised me a little bit.

When I first introduced the special forces unit into my mod it had a high attack value, bombard ability, blitz, all terrain as road, in other words it was a super unit, and the cost was not in line with the abilities. The ai built these like crazy and almost beat me with them.

Hi Wladislaw,

Not that I expect you to go through this entire thread (indeed, I'm going to try to summarize the results so far over the weekend) but the AI in no way seems to perform a "cost / benefit" analysis when deciding what to build -- mysterious Other Factors are at work here ...

@Rocoteh -- "Welcome back, my friend, to the show that never ends ..."

Best To All,

Oz

 

[The Last Conformist]

I've not done any research exactly on the subject of AI build choices since this thread last surfaced, but I did find out how drafting prioritization works. FYI:

i) To be able to draft at all, you must be able to build at least one unit with the Draft flag checked. (This might seem obvious, but plenty of people have refused to believe it. It's nonetheless true, and means you can't have units that can be drafted but not built.) If you only have one such unit available, it's what gets drafted.

ii) If you've got several such units available, the one with the highest defense factor gets drafted.

iii) If there is a tie for defense, the unit with the lower cost gets drafted.

iv) If cost is also tied, the unit coming first in the units list gets drafted.

I belive (iii) to be a bug - it ensures that Guerilla never gets drafted in the Epic game, despite having the Draft flag, which can hardly be intended. Should probably work the other way around - the more expensive unit should get drafted.

 

[Rocoteh]

The Last Conformist,

Interesting comment. What you mention was
unknown to me.

ozymandias,

Somewhat off-topic: I had planned to use the treasure-units
to simulate the Battle of the Atlantic.
Transports (with treasure units) from America to United Kingdom
attacked by U-boats. It does not work.
Do you have any knowledge on this subject?

Best Regards

Rocoteh

 

[10Seven]

Well, I did read the entire thread this morning - which must explain how it suddenly became 4:00 in the morning And now it's a little hard to see the point in bed, when one has to wake in an hour anyway

SUMMARY
For what it's worth, as this is my impression.

AI Build decision is based on.

1. FORCE COMPOSITION
- a ratio of offense versus defence. This appears to be the case, in consideration of the AI seeming to maximum produce certain units, in relation to losses. And in relation to the figures posted as to % of production devoted to unit types. And that the AI can only 'handle' a limited range of units in it's composition - perhaps 5 core units, with point 2. affecting any others.

2. RANDOM FACTOR

oftimes randomisation is added en leui of realism, and in explanation as to 'anomolies'. Even lacking evidence, I would assert this point, since it is so prevailant in PC games.

3. COST/A-D-M/SPECIAL ABILITIES

This is suggested by various posts as to AI changing it's build behaviour as a result of changing abilities - blitz, etc.

4. STATE OF DIPLOMACY
And state of war encouraging greater unit production over most buildings

With all these point revolving around each other, I think point 1. might be the highest of all.

ARTILLERY
I'm wondering if this can be tagged as requiring 'escort' thereby utilising the defensive bombard - together with ZOC, artillery would be far less powerful than current, but this reduced power would 'fit' with the AI utilisation.

It may appear that the AI was built, in relation to artillery, to be far less powerful than it is. Perhaps it was intended to work differently, but was changed due to popular demand?

TREASURE
There has been some discussion and testing undertaken by the Storm Over Europe team.

It appears to be the usual under-utilisation that tends to afflict anything related to water base activity. It was found, briefly, that there may be in the region of 80 treasures 'waiting' for transport, and only one transported, after 100 or so turns...

 

[Ozymandias]

Per conversation with Rocoteh, I've decided to do some more testing as to what decisions the AI makes in its builds.

So far, this exercise has (and shall continue to, at least for now, unless anybody wants to pick up the ball and run with it ) focused on land units.

I want to test flags, so I'm going to make 6 variations of the ol' familiar 5/5/1 A/D land combat unit, each with one of the following flags checked:

• Blitz
• Radar
• Amphibious
• Paradrop
• Invisible
• ZOC

My plan is to run 5 "round robins". With each round, I'll note the AI's most prevalent choice, then drop that choice for the next round; note the next most prevalent choice and drop it for the following round; etc.

Although I'll note proportions for each build choice, I'm doing this instead of simply having a "free-for-all" of one round of all six choices as I suspect the AI's choice will vary depending upon which is available ... Comparing the first round to subsequent rounds should verify or negate this.

Thoughts / insight / comments / questions before I begin?

Thanks,

Oz

PS I'm not testing the "All Terrain As Roads" flag as earlier testing showed that the AI was oblivious to MF.

-O.

 

[Ozymandias]

... Going once, going twice ...

Okay, unless anyone has any other input, I plan to run this test tomorrow / Saturday.

-Oz

 

[mrtn]

This sounds like a great idea Ozy. Looking forward to the result.