Will the AI ever get better at templates?

[Faeelin]

I know there's a new DLC coming out, but I have to ask: will there be any improvements to the AI coming? I think most players are aware of the ways the AI struggles with templates (making tons of ones it doesn't use, not upgrading templates as time goes on, etc.). Rather than a Finnish focus tree, this seems like something which would add value to all players.

 

[Leinad965]

I am not aware of AI struggle with templates. While EaI templates are definetelly better, vanilla templates are decent. AI is bad in production, variant creation, research, attacking, naval warfare, strategic bombing, ... But templates? Some players even crying in forums about loosing against powerfull AI templates.

 

[Voigt]

Overall the ai is decent with templates, but there should be more nation specific ones I feel.
China with pure infantry and no suppprt companies, beside maybe aa1.
France and Poland with 10/0 Infantry and some suppport arty/aa/support.
Germany/UK with maybe 7/2, since AI works better with more decent troops, than a mix of good and bad ones.

 

[jpd]

The AI doesn't upgrade templates by design. In other words, it's intentional. And it works from template design x to template design y through incremental steps, ie. chance one battalion and save. Presumably, this makes it easier to manage. And yes, that does mean it creates templates (the intermediate steps) that it will never use to create divisions with. But that's not a problem. The AI has no trouble figuring out which of these templates it will use to create divisions. It's only a problem for a human when a human player tag switches to an AI nation, and sees that long scroll list ...

Changing an existing template (at least when you only add battalions to it) creates an instant, big demand for additional equipment. Something a human can deal with quite easily (oh shit. I can't save now. It requires 10K art units, while I only have 2k. Need to save more, and try again later), but the AI has issues with. I haven't seen the AI save up equipment to date. And, without a saved stockpile of equipment, upgrading existing divisions would instantly weaken the AI, until production has caught up, and reinforced all those divisions.

 

[PurpulaPhoenixum53]

I agree with the sentiment above. The AI is really good a creating templates for the late game and using them. What the AI really needs to be improved on is:

1) Encirclement - The AI stalls out when it/I encircle some troops. This isn't helpful and that pocket should be taken care ASAP. This is most pronounced on the Axis, when they manage to pocket a major Allied landing, they end up staring at all those troops while not taking the initiative to pop the pocket

2) Garrisons during war - I still find the AI has some of its troops on garrison when its at war. Unless there is good reason to have troops on garrison, the AI should be taught to bring all available troops to the fronts.

3) Improve front maintenance - This pretty bad with the Axis again. The AI is constantly moving troops from front to front, allowing you at times to seize an initiative, but when they come back to stall out. I shouldn't be replaying WW1.

4) Troop spamming - The AI shouldn't be spamming troops into space at the height of the war. There should be some type of cap (perhaps based on conscription law and CIV factories) of possible number of troops. This should prevent nations like El Salvador from having 25 divisions, on two tiles. This should also help the USSR get an advantage against the Axis, as they will come late game begin to massively outnumber them.

5) Prevent Conscription Law "Leap-frogging" - This is more a base rule issue. The Soviets tend to jump to All-Adults serve too early into the game. Prevent the AI (and perhaps the player) from leap-frogging forward on conscription laws, by making them incremental and require you have low manpower before you change it.

6) Better inter-AI cooperation - The AI should be a bit better at working with its allies when comes to war. For example, if Britain launches the liberation of Greece with 45 divisions, Greece should try to invest into its infrastructure to help Britain maintain those divisions, while also sending its own divisions.

7) Infrastructure - Speaking of infrastructure, the AI should try and factor in infrastructure when it sends in an invasion. If it cannot handle the infrastructure, it will build more (if it can), or the ally will attempt to boost infrastructure.

 

[blahmaster6k]

I am not aware of AI struggle with templates. While EaI templates are definetelly better, vanilla templates are decent. AI is bad in production, variant creation, research, attacking, naval warfare, strategic bombing, ... But templates? Some players even crying in forums about loosing against powerfull AI templates.

I laugh every time I see people complain about bad AI and then also complain that they don't want to fight good AI division templates.

 

[blahmaster6k]

1) Encirclement - The AI stalls out when it/I encircle some troops. This isn't helpful and that pocket should be taken care ASAP. This is most pronounced on the Axis, when they manage to pocket a major Allied landing, they end up staring at all those troops while not taking the initiative to pop the pocket

This is a problem with the battle planner, not just the AI. The front line system behaves really oddly when pockets are created, assigning some random number of divisions to a new front surrounding the pocket, taking away from the main front, causing troops on the main front to shuffle back and forth to try to spread themselves out evenly, and pausing the offensive while this happens. Likewise, the troops automatically assigned to surround the pocket won't attack it because there's no attack order pointing into the pocket (unless set to aggressive, which makes them attack without an arrow drawn if the army is told to activate plans.

Garrisons during war - I still find the AI has some of its troops on garrison when its at war. Unless there is good reason to have troops on garrison, the AI should be taught to bring all available troops to the fronts.

This is intended to counter players just sniping all VPs with paratroopers and instantly cheesing entire wars, which was a major strategy a long time ago.

 

[SchwarzKatze]

The AI tries to produce whatever template it's scripted with. Most countries have 7/2 templates, but some majors have "historical" templates that just don't work well. For example, Germany is made to produce 27 width infantry while Japan is scripted to produce 19 and 26 width infantry templates.

 

[kettyo]

Changing an existing template (at least when you only add battalions to it) creates an instant, big demand for additional equipment

Not so much if you change as many divisions as you have stockpiled equipment for. I often find myself changing one by one as equipment is produced, especially with new stuff and low efficiency. The AI should be able to manage this too. The part where it'll have major difficulties is deciding where the new templates are needed most.

 

[Michael Gladius]

Here are two ways to improve them:

Divisional composition should affect tactical decisions

Right now, a 40-width Division of 28,000 men will steamroll over a more historical 11,000-man division, and be able to sustain its momentum near-indefinitely. IRL, the massive, lumbering divisions of WWI were quickly dropped in favor of smaller...

forum.paradoxplaza.com

https://forum.paradoxplaza.com/forum/threads/land-doctrine-upgrades.1379687/

 

[Timmysoboy]

Gladius: I like the first option (tactics), not big on the second. Also, I’m really hesitant to ask for corps commands, unless it is stupidly user friendly. Less is sometimes more in terms of game enjoyability for me.

#5 from purpula: looking at the AI play, I think the issue may be that they’re boosting their conscription law whenever they’re at low manpower, regardless of how much more manpower can be gained under their current law. For example: they switch from extensive to service whenever they hit 0, even if they’re rocking 2.6% recruitable.

On topic: I wouldn’t hate it if the AI deleted unused templates.

 

[bitmode]

Not so much if you change as many divisions as you have stockpiled equipment for. I often find myself changing one by one as equipment is produced, especially with new stuff and low efficiency. The AI should be able to manage this too. The part where it'll have major difficulties is deciding where the new templates are needed most.

The AI might be able to but its current system is (in this regard) strictly superior.
There is no downside to having many templates and the upside of saving XP when branching in different directions. There is no need to combine template editing and unit upgrading into one action.

 

[el nora]

The AI might be able to but its current system is (in this regard) strictly superior.
There is no downside to having many templates and the upside of saving XP when branching in different directions. There is no need to combine template editing and unit upgrading into one action.

There is one downside, and its a big one imo. The loss in veterancy. Converting from one template to another will lose veterancy, even if the new template has a lower manpower requirement. Editing a template down to another with lower manpower requirements does not lose veterancy.

For example, If my 20 infantry grinding bricks that I send to the SCW become veterans, I could swap them over to my 15-5 template, but they would drop down to regulars. If, instead, I edit their template to 15-5, they keep their veteran status with its +75% stats.

 

[bitmode]

Editing a template down to another with lower manpower requirements does not lose veterancy.

I consider this to be a bug.

 

[el nora]

It's been nearly 3 years. Since 1.4. It may be a bug, but not one that PDX seem keen on fixing.

 

[kettyo]

The AI might be able to but its current system is (in this regard) strictly superior.
There is no downside to having many templates and the upside of saving XP when branching in different directions. There is no need to combine template editing and unit upgrading into one action.

That's correct but we've talked about a different thing.

I've only argued that the AI can upgrade units one by one just like a human would do and not forced to stockpile an insane amount of equipment to upgrade the complete division type at once.

To put it in context let's say it has 100 divisions of infantry and starts building anti-tank to introduce support anti-tank into them. It does not have to stockpile 2400 anti-tank to upgrade all at once but can upgrade each unit as it has a spare equipment of 24 one by one which is much more logical and optimal as equipment in the stockpile provides no befenits.

 

[bitmode]

That's correct but we've talked about a different thing.

I've only argued that the AI can upgrade units one by one just like a human would do and not forced to stockpile an insane amount of equipment to upgrade the complete division type at once.

To put it in context let's say it has 100 divisions of infantry and starts building anti-tank to introduce support anti-tank into them. It does not have to stockpile 2400 anti-tank to upgrade all at once but can upgrade each unit as it has a spare equipment of 24 one by one which is much more logical and optimal as equipment in the stockpile provides no befenits.

Oh I thought the AI already does it incrementally, which is why I got confused.

Edit: there are twothree AI defines indicating incremental unit upgrades:

    UPGRADE_DIVISION_RELUCTANCE = 7,                    -- How often to consider upgrading to new templates for units in the field
    UPGRADE_PERCENTAGE_OF_FORCES = 0.1,                    -- How big part of the army that should be considered for upgrading
UPGRADES_DEFICIT_LIMIT_DAYS = 50,                           -- Ai will avoid upgrading units in the field to new templates if it takes longer than this to fullfill their equipment need

 

[kettyo]

Oh I thought the AI already does it incrementally, which is why I got confused.

Edit: there are twothree AI defines indicating incremental unit upgrades:

    UPGRADE_DIVISION_RELUCTANCE = 7,                    -- How often to consider upgrading to new templates for units in the field
    UPGRADE_PERCENTAGE_OF_FORCES = 0.1,                    -- How big part of the army that should be considered for upgrading
UPGRADES_DEFICIT_LIMIT_DAYS = 50,                           -- Ai will avoid upgrading units in the field to new templates if it takes longer than this to fullfill their equipment need

I supposed it's done so but someone before me had the idea that the AI is bad at implementing templates because upgrading a complete division type needs tons of equipment. Which is true but it's not much relevant.

 

[bitmode]

I did some digging in the AI code. For each existing template, the AI first checks which other template would be the best upgrade according to one of its roles. Only then does it start incrementally upgrading fielded divisions according to these choices. There are two modes:

1. Finding a closer match to the ai_template the template already best fits to. E.g. a generic minor given an infantry template looks for another template that more closely matches the 7/2+ENG,REC,HOSP,AT,ART target template
2. Upgrading along a replacement path. E.g. when Germany upgrades from light to medium tanks due to the replace_with = medium_armor_default_GER attribute in their light tank template.

But it doesn't work in most cases because these approaches can't be mixed:

Finding best existing upgrade for "Infantry Division"
role entry "infantry_generic" has matching role "infantry"
best matching AI template in role entry is "infantry_default" with score 0.358
can_upgrade_in_field is true but replace_with is unset! skipping

Spoiler: gdb script for 1.9.3 linux

break *0xfce790
commands
silent
printf "Finding best existing upgrade for \"%s\"\n", *(char**)($rsi+0x28)
c
end
break *0xfce8b0
commands
silent
printf "role entry \"%s\" has matching role 0x%x\n", *(char**)($r9+0x10), $esi
c
end
break *0xfce927
commands
silent
printf "best matching AI template in role entry is \"%s\" with score %.3f\n", *(char**)($r12+0x10), $rbx/1000.
c
end
break *0xfce967
commands
silent
printf "both can_upgrade_in_field is true and replace_with is set in best matching AI template, finding better successor\n"
c
end
break *0xfce824
commands
silent
printf "can_upgrade_in_field is true but replace_with is unset! skipping\n"
c
end
break *0xfce9ae
commands
silent
printf "can_upgrade_in_field is false, looking for closest match to AI template\n"
c
end
break *0xfcea53
commands
silent
printf "Template successor is \"%s\"\n", *(char**)($rax+0x28)
c
end
break *0xfcea27
commands
silent
printf "Closest template is \"%s\"\n", *(char**)($rax+0x28)
c
end

Except for a few light tank templates, no AI template definitions have can_upgrade_in_field, which is true by default. This means even when you give an AI country artillery for free and add support artillery to a copy of their starting template, they will not incrementally start using it.
Eventually after a while they switch over the entire army to the new template using some other mechanism, but not with the incremental defines mentioned earlier.

When adding

 can_upgrade_in_field = {
  always = no
}

to infantry_default in infantry_generic in common/ai_templates/generic.txt, the AI is able to pick up the new template:

Finding best existing upgrade for "Infantry Division"
role entry "infantry_generic" has matching role "infantry"
best matching AI template in role entry is "infantry_default" with score 0.358
can_upgrade_in_field is false, looking for closest match to AI template
Closest template is "With support artillery"

and switches its divisions over soon after in small batches as equipment and manpower availability allows. I believe all AI templates without replace_with should be changed like this.

@stjern

 

[kettyo]

@bitmode I'm not quite sure if i understand all this but the German template file has a funny bug. light_armor_anti_amor_GER can replace_with = medium_armor_anti_armor_SOV. The next generation of German anti-tank is the Soviet anti-tank