Unit controllers want just one thing and it's f**king disgusting

[bitmode]

Digging through the code responsible for executing battle plans is like the proverbial yak shaving. But one thing I've already noticed is that it throws almost everything it does out the window in the final step: spreading attacks.

This is how the Italian AI draws its first offensive line in northern Ethiopia; let's assume this is a reasonable plan. To faithfully execute it, the unit controller correctly identifies that the northern-most mountain province is most important. It is among the furthest away from the offensive line, capturing it does not lengthen the frontline and it can be attacked from three directions. Nice!
Then the shooting starts:

Four divisions attack the central province, three the western one, and one is still in transit. As soon as some division target the central mountain province, its score tanks -- alot. A single division targeting a province divides its score by 13*, two divisions by 26 etc. After two or three divisions the remaining scores are zero and further assignments are essentially random. I think this is why AI attacks feel like a horde of orcs smashing against Helm's Deep.

* 00_defines.lua: NMilitary.PLAN_SPREAD_ATTACK_WEIGHT

 

[Alex_brunius]

Did you test reducing the defines to see how it impacts AI behavior?

 

[bitmode]

Did you test reducing the defines to see how it impacts AI behavior?

Yes briefly, the define can be lowered down to 0.5 (below that will cause a division by zero crash). It makes the attacks much more focused; in the Ethiopia example all divisions attack the middle province first.
Afaict the unit controller does not consider combat width though. So this change in isolation would surely lead to over commitment on more saturated fronts.

 

[Alex_brunius]

Afaict the unit controller does not consider combat width though. So this change in isolation would surely lead to over commitment on more saturated fronts.

Over commitment is not necessarily only bad. It's an exploit used by players to farm leader traits, so getting the AI to do the same, and keep attacking in the same place might even be desirable. Your right that it's a risk if it's done in bad conditions though like when suffering attrition through attacking mountains or such.

Now we just need to figure out a way to get the AI to not constantly swap around it's Generals and fill armies up better to 24 divisions before making new ones. Got any ideas?

 

[grandad1982]

People couldn't stop staring at this fallback line and it didn't know why.....

 

[EntropyAvatar]

A single division targeting a province divides its score by 13*, two divisions by 26 etc. After two or three divisions the remaining scores are zero and further assignments are essentially random. I think this is why AI attacks feel like a horde of orcs smashing against Helm's deep.

Where do people come up with this stuff? You would think the default would be reduce score by fraction of combat width already committed, and modify from there.

 

[bitmode]

Where do people come up with this stuff?

In Sweden

Now we just need to figure out a way to get the AI to not constantly swap around it's Generals and fill armies up better to 24 divisions before making new ones. Got any ideas?

Unless a define is laying around that smells like it could do that, probably not.

 

[Reman]

Interesting how the mechanics work. Is spreading the attacks a bad thing in your view? In an average situation I'd probably prefer it. Using org sponges to lock adjacent units in combat tends to be the only way to reliably take provinces in offensively difficult situations. Slamming into a single province over and over is ineffective when it can be refreshed with units from adjacent provinces. I've seen the battleplanner play both sides of this to some effect.

 

[bitmode]

Interesting how the mechanics work. Is spreading the attacks a bad thing in your view? In an average situation I'd probably prefer it. Using org sponges to lock adjacent units in combat tends to be the only way to reliably take provinces in offensively difficult situations. Slamming into a single province over and over is ineffective when it can be refreshed with units from adjacent provinces. I've seen the battleplanner play both sides of this to some effect.

Spreading is not bad in itself. It depends on how the forces match up and what the higher level goal of the attack is (grinding down enemies, capturing a VP etc.) If the primary battle is not green to begin with - which is often the case from the AIs perspective - pinning is not so important.
Not all units need to be committed to a battle in the first hour. This is especially noticable in careful plan setting. Doing a primary attack first and then adding pinning attacks around it if it's going well is something I frequently do. This should be possible for the unit controller with its current data and complexity. (it already has a bunch of logic about which existing battles to support like RESERVE_TO_COMMITTED_BALANCE)

Generally I rather have an AI that does something wrong for the right reasons than doing something right randomly, but that's personal preference.

 

[bitmode]

Also something that annoys me in this game: the attacks are not truly random but "implementation defined" i.e. whichever enemy province is the "first" neighbor by internal ordering gets selected. If the province IDs (the ones you see with tdebug) were assigned systematically across the map, this behavior becomes quite obvious because divisions would tend to attack in a certain orientation. Randomizing the province IDs makes this harder to detect without actually solving the problem and probably hurts performance too.

Edit: in the Ethiopia example the western province (1984) has the smallest ID

 

[Louella]

the define can be lowered down to 0.5 (below that will cause a division by zero crash)

Does that not imply that it's rounded off, so there'd be no difference between setting it to say 6.6 and 7 ?

Generally I rather have an AI that does something wrong for the right reasons than doing something right randomly, but that's personal preference.

Yeah, like, "attack this province with these units because the terrain is favourable and opposing forces are weak to the attacking force." is something that might be right, or might be wrong, depending on how it evaluates the terrain and strength of enemy. But that's something that can be worked with, by twiddling how it evaluates things.
"attack this province with these units because apples are superior to oranges", what even is happening here ? Sometimes it might work, but, you have no idea why, or how you could twiddle things to make it work more often.

 

[billcorr]

As of this posting, the OP has been rated one of the most helpful (non-dev) posts in a while.

Wonder if it will get 50 "helpfuls"

Nice job!

 

[Achab]

As of this posting, the OP has been rated one of the most helpful (non-dev) posts in a while.

Wonder if it will get 50 "helpfuls"

Nice job!

View attachment 444830

What a shameless popularity contest bump!

 

[bitmode]

Does that not imply that it's rounded off, so there'd be no difference between setting it to say 6.6 and 7 ?

It's multiplied by number of attackers first and then rounded, so both 1 and 2 attackers would be rounded to the minimum divisor 1 when setting the define to 0.5

Having tested 0.5 in a longer game, it's not an improvement by itself because the attacker uses normal move commands and leaves the provinces they came from undefended after concentrating an attack. (Even more so than they do with vanilla settings). Since patch 1.5 the defenders are very good in exploiting such openings and create encirclements out of that.

 

[Louella]

Having tested 0.5 in a longer game, it's not an improvement by itself because the attacker uses normal move commands and leaves the provinces they came from undefended after concentrating an attack. (Even more so than they do with vanilla settings).

Hmm. Is there something that would get the attacker to use the "support attack" thing more often ? The thing which makes them join a battle, but doesn't make them advance when the battle is won ?

 

[bitmode]

Hmm. Is there something that would get the attacker to use the "support attack" thing more often ? The thing which makes them join a battle, but doesn't make them advance when the battle is won ?

Yes, in principle. On careful plan execution it is used quite frequently. It even looks like the unit controller can initiate support attacks as the battle starts. I haven't reversed all the details yet but RESERVE_TO_COMMITTED_BALANCE and PLAN_MAX_PROGRESS_TO_JOIN are the two defines that should have at least some influence.
The former just balances how many divisions can sit in the reserve of a battle compared to how many are actively fighting (10% seems a bit low).
The latter looks to be very broken. "Progress" seems to refer to the sum of organization of all divisions. Only if that sum is below the threshold (0.5) will certain support attacks be executed. Judging from the code, setting this to one million is a good course of action...
In any case both of these are almost orthogonal to the problem of vacating provinces.

 

[bERt0r]

Yes briefly, the define can be lowered down to 0.5 (below that will cause a division by zero crash). It makes the attacks much more focused; in the Ethiopia example all divisions attack the middle province first.
Afaict the unit controller does not consider combat width though. So this change in isolation would surely lead to over commitment on more saturated fronts.

But this is like literally everything I ever wanted from the battlep lan system! If I want a second offensive going on I'm gonna draw a second battle plan. Why lower the importance of a province for each unit that attacks it? Make that value 1 and be done with it. This is like the dumbest thing ever.

 

[kaguravitro]

Spreading is not bad in itself. It depends on how the forces match up and what the higher level goal of the attack is (grinding down enemies, capturing a VP etc.) If the primary battle is not green to begin with - which is often the case from the AIs perspective - pinning is not so important.
Not all units need to be committed to a battle in the first hour. This is especially noticable in careful plan setting. Doing a primary attack first and then adding pinning attacks around it if it's going well is something I frequently do. This should be possible for the unit controller with its current data and complexity. (it already has a bunch of logic about which existing battles to support like RESERVE_TO_COMMITTED_BALANCE)

Generally I rather have an AI that does something wrong for the right reasons than doing something right randomly, but that's personal preference.

Did you rise org and str values in frontline logic define lines, I use some similar to expert Ai and general situations have better development, not a solution but accord weak attacks

Actually I'm most interested in advance and reorganization task of front logic, I tested a lot and looking fine o I found that setting generals to 15 units and setting width 60+x20 frontline advance is better, defender avoid pockets from inf (armor in Ai vs Ai still could make pokets), makes line enough reactive in offender and defense.

But even with this commendation of the issue longest disgrace to rebalance front, still there are problem, advance doesn't difference terrain no matter values used in define and as we doesn't did something as an attack that different move undid to attacked province like we have in hoi2, you make deisbalance in every advance advanthe line.

 

[bitmode]

Make that value 1 and be done with it

As I said, that's not enough by itself because the attackers bunch up too much and get encircled.

This is like the dumbest thing ever.

Please keep it constructive; I don't want to get the thread locked by the moderators.

 

[Louella]

the problem of vacating provinces.

I remembered a define that I saw in the past:

NAI section, first entry:
GARRISON_FRACTION = 0.0, -- How large part of a front should always be holding the line rather than advancing at the enemy

Maybe this bit is why the AI has a tendency to vacate provinces ?