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

[The Last Conformist]

Should be interesting.

I'd throw in an ATAR unit anyway. I would not be entirely surprised to learn it puts a value on the flag, regardless of ignoring MF.

 

[Rocoteh]

ozymandias,

I have been away for some days, thus the silence.

I am very glad that you have reactivated this thread.
It will be interesting to see the results of the experiment.
Invisible:
I think AI will like that....

An observation from playtesting WWI-Global1.9c:
By, mistake an powerful marine-unit with very high shield cost
become available early in the game:
Russia AI shifted production to produce 80% Marine units!
Thus Russia was knocked out early, since it produced units that
in most cases took 25-30 turns to complete.

Rocoteh

 

[Ozymandias]

I wound up running a series of tests qwith two AI Civs each with 5 cities, running 5 turns each test for a total of 50 AI builds (as seen earlier -- if you can remember that far back in this thread -- the AI Player Governer builds differently than for AI Civs; go figure).

I goofed in the first go-round, with Amphibious and Paradrop units unavailable (I forgot to uncheck the resource requirement).

1.So the first round tested Blitz, Radar, Invisible, ZOC, and All Terrain As Roads (ATR).
Observation In this test, one Civ built ONLY Radar on turn one, whereas the other built only ATR ...
Conclusion I'm confused ...

Result: first turn EVERY city built Radar; thereafter production was 100% Invisible!
Conclusion: We've seen earlier that the AI likers to build a defensive unit on the first turn (even when I re-ran this test with Amphib and Para included, the results were the same) and therefore "recognizes" Radar as the best defensive flag for an initial build; thereafter Invisible reigns supreme for both offense and defense.

2. Eliminating Invisible from the pool, 38 out of 50 builds were ATR (thank you, The Last Conformist!)

3. Eliminating ATR from the pool showed an evenly divided preference for ZOC and Blitz.

4. I then ran some less rigorous tests (trust me, this stuff is tedious) indicating a preference in the following order: Radar, followed equally by Para and Amphib.

Overall Conclusion: yes, folks, it's all relative; although preferences seem established, ABSOLUTE preferences are not. There's a randomness factor in here somewhere.

-- My next effort will be to see if the AI (all things being equal) prefers to build a unit which requires a special resource or one which does not. Stay tuned ...

-Oz

 

[Neomega]

I wound up running a series of tests qwith two AI Civs each with 5 cities, running 5 turns each test for a total of 50 AI builds (as seen earlier -- if you can remember that far back in this thread -- the AI Player Governer builds differently than for AI Civs; go figure).

I goofed in the first go-round, with Amphibious and Paradrop units unavailable (I forgot to uncheck the resource requirement).

1.So the first round tested Blitz, Radar, Invisible, ZOC, and All Terrain As Roads (ATR).

Result: first turn EVERY city built Radar; thereafter production was 100% Invisible!
Conclusion: We've seen earlier that the AI likers to build a defensive unit on the first turn (even when I re-ran this test with Amphib and Para included, the results were the same) and therefore "recognizes" Radar as the best defensive flag for an initial build; thereafter Invisible reigns supreme for both offense and defense.

2. Eliminating Invisible from the pool, 38 out of 50 builds were ATR (thank you, The Last Conformist!)

3. Eliminating ATR from the pool showed an evenly divided preference for ZOC and Blitz.

4. I then ran some less rigorous tests (trust me, this stuff is tedious) indicating a preference in the following order: Radar, followed equally by Para and Amphib.

-- My next effort will be to see if the AI (all things being equal) prefers to build a unit which requires a special resource or one which does not. Stay tuned ...

-Oz

Nice work, very informative.

 

[Ozymandias]

Nice work, very informative.

Thank you -- FYI I edited it while you were posting (ah, cross-posting ...) you might want to glance again.

Best,

Oz

 

[Rocoteh]

ozymandias,

Very interesting!

To me, its clear that AI like "overprized" units per se,
though its hard to show a proof.

Rocoteh

 

[The Last Conformist]

That the AI is madly in love with Invisible is "old news".

Otherwise, the prioritization seems pretty sensible; ATAR is quite powerful (on offensive units), while things like Radar and Para aren't that great, in most circumstances.

It would be interesting to know if the preference for Amph is dependent on whether the AI's at war with an overseas enemy. At least back in PTW, the AI seemed to build warships only if it had an overseas enemy.

 

[Rocoteh]

With regard to this aspect I think that old news
confirmed also are interesting.

From my own playtesting today, I find it interesting
that small Civs like Norway (3 cities on an earth-map)
produces a high cost naval unit like a Dreadnought within 15 turns.

Edit: Above refers to WWI-Global scenario earlier mentioned.

Rocoteh

 

[Ozymandias]

I ran the test with just two units, a "Regular" 5/5/1 etc. and a "Strategic Resource Dependent" ("SRD") equivalent. The necessary strategic resource was present in abundance.

Build 1: 10 Regulars

Build 2: 10 Regulars

Build 3: 5 Regulars, 5 SRDs

Build 4: 7 Regulars, 3 SRDs

Build 5: 6 Regulars, 4 SRDs

-- I have no idea how to interpret this ...

-Oz

 

[Rocoteh]

ozymandias,

I think the "solution" is that its 100% random.
(After build 2)

I made your experiment, but this time both units "A" and "B"
were regular 5-5-1 etc. Thus no "SRD".

The result:

Build 1: Ax10, Bx0

Build 2: Ax9, Bx1

Build 3: Ax3, Bx7

Build 4: Ax2, Bx8

Build 5: Ax6, Bx4



Rocoteh

 

[mrtn]

ozy, could you make another test with 5/5/2 units? Having blitz on a one-move unit should (but probably isn't...) be recognised by the AI as useless...

 

[Ozymandias]

ozy, could you make another test with 5/5/2 units? Having blitz on a one-move unit should (but probably isn't...) be recognised by the AI as useless...

Sure. What shall I test it against? (hint: the less the number of other unit types, the easier the test ). BTW earlier results did show that the AI was "blind" to MF; that the AI built any Blitz units at all at MF=1 seems to support that.

-Oz

 

[Ozymandias]

ozymandias,

I think the "solution" is that its 100% random.
(After build 2)

I made your experiment, but this time both units "A" and "B"
were regular 5-5-1 etc. Thus no "SRD".

The result:

Build 1: Ax10, Bx0

Build 2: Ax9, Bx1

Build 3: Ax3, Bx7

Build 4: Ax2, Bx8

Build 5: Ax6, Bx4

Rocoteh

So it couldn't / didn't distinguish between two IDENTICAL units at all -- except to arbitrarily (?) pick "A" as the better defensive unit?

-- BTW, excellent control test!

Best Regards,

Oz

 

[mrtn]

Sure. What shall I test it against? (hint: the less the number of other unit types, the easier the test ). BTW earlier results did show that the AI was "blind" to MF; that the AI built any Blitz units at all at MF=1 seems to support that.

-Oz

I meant make all units 5/5/2, regardless of the flag you give them... Use the flags that seems appropriate (i e you can probably skip the "invisible" units, as those were prefered by the AI...)

 

[Ozymandias]

Can do

BTW, I unscientifically ran some trials varying the AI's aggressiveness -- For the full tests reported above, I set all the AI's equal across the board (aggressiveness; attributes; etc.) with aggressiveness = 3. Then I set it = 5 and ran a few rounds -- NO SIGNIFICANT DIFFERENCE WAS FOUND.

Tentative Conclusion The AI's Agressiveness does not effect what it builds!

-Oz

 

[Rocoteh]

So it couldn't / didn't distinguish between two IDENTICAL units at all -- except to arbitrarily (?) pick "A" as the better defensive unit?

-- BTW, excellent control test!

Best Regards,

Oz

ozymandias,

Thank you.
Yes, since A and B were identical in all aspects (except name),
I interprete the choices that AI made as random.

Best Regards

Rocoteh

 

[The Last Conformist]

With regard to this aspect I think that old news
confirmed also are interesting.

I didn't mean to imply it was uninteresting, tho I realize my comment could be read that way. I rather think it's faintly reassuring that there's sufficient consistency in AI builds that there can be old news.

So it couldn't / didn't distinguish between two IDENTICAL units at all -- except to arbitrarily (?) pick "A" as the better defensive unit?

I don't believe it picked "randomly". My tests of about a year ago indicated that the AI will always build the cheapest available defender in an empty city; I suspect that if there's a tie, it simply picks the one coming first in the units list (this would be consistent with the ultimate tiebreaker for drafting).

 

[Ozymandias]

I don't believe it picked "randomly". My tests of about a year ago indicated that the AI will always build the cheapest available defender in an empty city; I suspect that if there's a tie, it simply picks the one coming first in the units list (this would be consistent with the ultimate tiebreaker for drafting).

Ah-hah! -- I wasn't aware of that. Although, looking at both Rocoteh's and my tests, this would only seem to apply to the very first build or two in the city ... ?

-Oz

 

[Dom Pedro II]

I have such admiration for this Civ testing... to devote so much time to the game to run these scientific experiments... just wow.

 

[Ozymandias]

I have such admiration for this Civ testing... to devote so much time to the game to run these scientific experiments... just wow.

It's just to compensate for not being a brilliant unit creator