Game AI stuff

[nikolai1962]

This is for stuff not directly related to path-finding.

Used alpaca's stripped down map to make a mini testing map.




The map is set up in such a way as to not have any path-finding problems so i could test other things. Plan is to set up each faction one at a time with one region and change various values to see what moddable values make them expand faster against those rebel settlements.

Some initial observations.

1. I always assumed (with rtw) the only difference with the AI at different difficulty settings was the amount of cash it recieved. Seems to not be so. On medium the AI is much less aggressive even if you give it a lot of cash via king's purse. It doesn't seem to recruit as many units as it does on VH or target settlements as aggressively. The AI still expands, just slowly. On the plus side it doesn't run out of cash like it does on VH.

2. On H and VH the AI will unit spam as much as it can. It has a tendency to overdo this and end up with no income. This means if it isn't constantly expanding or the economy isn't set up to continually increase the cash flow into the game (high pop growth etc) then it gradually slips behind with teching up it's cities especially later on when the new buildings get very expensive.

3. Siege units suck badly in auto-calc. So that means they are bad for the AI to over-produce in every way possible (I thought the high attack might mean they were at least useful to the AI in auto-calc but they're not).

4. The "best unit" council mission rewards can cripple a small faction with limited cash as 4 knights = 1000 upkeep. Making them all cash rewards is better imo as it reduces the number of times the AI will stall through lack of cash. (You can change this in descr_mission.txt.)

5. Family member upkeep is too high for small factions. 200 for the general and 250 for the bodyguard can also economically cripple a small faction if they get too many early on in the game. (If you have recruitable generals you could make the bodyguard a free upkeep unit which would be a neat player/AI thing as the AI likes to keep generals in the cities whereas players like to have them in command. Otherwise just reduce the upkeep of either/both the family member wage and/or the bodyguard upkeep.)

6. Even on my testmap the AI can still get stalled attacking a citiy after a retreat. (On a convex region map it can take quite a lot of retreats before it stalls though.) So MTW2 still needs an AI movement bonus trait that triggers for AI generals that end their turn outside a settlement with 100% movement points. It wants to go up to 100% bonus depending on the scale of your map and your base movement points) and have an anti-trait that triggers when they end their turn in a settlement.

This way if a general is stuck after a retreat eventually the movement bonus will increase to the point where they are within one turn of their target and they'll start moving again. Won't fix all the problems but will reduce it a bit. need an anti-trait to clear the movement bonus again so it can re-trigger if they get stuck later on.

7. The unit recruitment settings can have a massive impact on how well the AI expands. If the available recruitment allows the AI to build lots of crap units then they expand much slower. Constraining the unit choices can lead both to more fun tactical battles (which was the thing i was originally thinking about) and to better autocalc expansion against rebels etc.

This is the kind of french army i want to see in a battle :)



Can't stress enough how big an effect changing the recruitment limits/refresh rates can have on the AI expansion. With one setup the french AI on VH could take all the rebel regions in 30 turns, with a slightly different early recruitment they'd barely get 4 or 5 regions within 30 turns. Removing obsolete units from the recruitment pools as buildings upgraded, reducing the amount of militia buildable in towns (for the french anyway) and having knights easy to recruit was best for the early french.

The key thing is whether the units they can build are capable of beating the rebel garrisons or not. If they can't then the rebels keep getting experience upgrades (which seem to have a big effect in autocalc) so the AI keeps retreating over and over until they get stuck. In the screenshot above it took about 90 turns to get all the rebs as opposed to 30-ish when i had higher levels of knight recruitment.

8. Revolts and brigand activity can be quite useful as a cleaning mechanism for an AI faction that has run out of cash or has stalled some other way and also to remove obsolete armies but it depends a lot on difficulty level. On VH a higher level of brigands/revolts can be indirectly helpful to the AI whereas on M it is pretty much all negative. Seems to me it is best to mod for a specifc difficulty level and always play test at that level to get the balances right.

9. Knight dominated armies are great for the french faction. Move fast, can't assault so they siege out a city instead then cream the garrison in a sally auto calc. Great stuff.

10. In rtw, when the AI had no targets they'd concentrate round the capital. Now they congregate somewhere more sensible, either a central spot or near an enemy even if they can't attack (I have the AI set to only attack rebels currently but thye group around my border anyway after all the reb regions are gone.)
~~~

Will add more as i spot new things.

[SigniferOne]

Nikolai, very very good investigation. It will come extremely useful I think, to everyone who reads this thread. Here I brought your attention to an aspect of the campaign map, that might be worth investigating in conjunction with all this.

[nikolai1962]

To expand a bit on the unit recruitment point. As a baseline for testing i started with all units having a recruitment of 0.5 2 i.e max pool of 2 with a replenish rate of 0.5. It wasn't that the french could produce lots of knights (as they didn't have the cash at the start) but the proportion of good units to town militia. As the vanilla numbers for town militia are something like 0.9 8 the french AI could spam them each turn while they couldn't spam the better units. So for the same amount of income/upkeep their army ended up with 70%+ town milita and ballistas. When i'd changed all the pools to 0.5 2 they used that same amount of income to produce a smaller total number of units but with a higher percentage of good units.

[nikolai1962]

Back looking at MTW2 modding again and thought i'd mention a few other things I'd noticed. After this amount of time I guess most modders know this kind of thing but just in case there are still newbie TW modders.

An example of point (5) in the original post about the upkeep cost of generals. Venice in the vanilla game has four start generals and also four kids in the family tree ready to come of age (compared with say England who has three and one respectively). This means they'll be paying for eight generals quite early on in the game which is roughly 4000 florins which they usually can't afford as they don't have an easy early expansion.

Obviously generals are a big advantage as well as a cost, but if you mod things like the 2HP of the generals or reduce king's purse and then wonder why Venice doesn't seem to do very well (as an AI faction) this is part of the reason.

~~~

This brings me on to generals and autocalc. Autocalc is heavily influenced by:

1) Number of soldiers - even if they're crap
2) Melee attack rating

(why horse archer AI factions were so useless in RTW without lots of infantry or sneaky modding)

Autocalc is also very heavily influenced by multi-hitpoint units and command stars.

(One of the reasons the Egypt and Britons AI factions always did so well in Vanilla RTW was because of this--2HP as a general plus a bunch more for the bodyguard being chariots.)

What this means is, if the general unit has very high stats, multi-hitpoints and an AI command bonus, then to the game it is a very strong auto-calc "stack", even on its own. Given that the AI likes to leave a general in each city it owns, and only leads armies with spare generals, this can have the effect of the AI often wandering round fighting with just a general (and maybe a few scraps of damaged units). You can see this in the vanilla game with the byzies starting 10 star general.

Very high stats and multi-HP is also partly why the generals in TW games suicide charge. They do a one-on-one match up with the player units nearby and think "charge"--not taking acount of the other twelve player units waiting to jump in. Weaker generals are (counter-intuitively) better. (Also why high anti-cav bonuses are bad imo as it makes spear units break formation to chase cavalry.)

Generally I think player-only handicaps work better than AI bonuses in TW games because of unintended consequences like this.

Also, something that only occurred to me the other day playing a mod and seiging another general-only AI garrison, is *if* the general on its own is a very strong autocalc stack, then to the game it might meet the AI's requirement for a garrison. As mentioned the AI always likes to leave one general per city if it has them and if the general unit is very strong that might increase the odds of the general unit being the *only* garrison unit.

Haven't tested that last point yet. Reduced the stats and HP on the generals in my modding for the other reasons mentioned and the AI garrisons have been a bit bigger but it may be due to some other factor.

I'm going to use the idea in conjunction with something I saw in the UAI mod which I think has potential which was to increase base unrest and make the garrison effect a bigger proportion of unrest control.

(Another aspect of this potential general effect vis a vis garrisons is that the strong stats/multi hit points of the general make it a strong autocalc unit but doesn't help much with unrest because of the low numbers in the unit. Hence possibly why the AI loses remote regions to revolt so often after taking them. They leave the general behind as garrison and move out.)

~~~

Which brings me to revolts. One of the problems with the Moors in the vanilla game is they go after timbuctoo and keep losing it, so they have to keep re-taking it. (You see this stuff if you watch each AI faction in turn with fow off to see what their problems are.) Moorish armies heading for Timbuctoo, either the first time or all the subsequent times, also rebel a lot because of the distance from capital etc so it adds up to a big distraction for the Moorish AI very similar to what happened with Carthage in RTW. A simple change like making timbuctoo a castle with a lower pop stops this problem. They take it and don't lose it making them expand better.

~~~

Which brings me to Spain. Best way to get AI factions to fight is to rig the start position so two factions go for the same rebel region(s). On the vanilla map (partly cos of map path-finding problems) the Moors concentrate more on Africa, and Spain/Portugal head north into France, or even more annoyingly, Ireland ffs. Changing things round so there is a belt of closer rebel regions in between the Moors and Portugal/Spain makes for a proper reconqusita war almost every game, which makes playing one of those factions much more fun.

~~~

Turks. Number of things hold up the vanilla turks. Path-finding is one but also semi-related to that is the lack of roads in their area which can be easily fixed with a few map tweaks. This reduces their travelling time to attack targets which matters a lot for them because their starting capital is set to iconium but most of their early rebel targets are in the opposite direction. This means their early attacking stacks rebel *a lot*. This is exacerbated by their region having a lot of narrow mountaim passes so the rebelled stacks then block future armies moving to attack the same city.

A good fix for the turks is making a new region in the centre of their start area and making that the capital. This reduces the early rebellions as the "distance to capital" rebellion effect is reduced. Adding some authority traits helps with this as well, good for all factions that have armies moving a long way from their capital to attack rebel settlements in the early game(Russia also suffers a lot from this).

Making the roads work reduces the number of turns they spend travelling to their targets so that reduces the rebellion chance also. Their early path-finding problems are mostly fixed with higher base movement points.

~~~

Milan. One of the reasons Milan is so strong in vanilla is the unit upkeep of their militia is low.

~~~

France: Some pathing problems but also the relative proportion of castles/towns in France doesn't suit their strengths (suits Milan much better, hence them beating France so often in the vanilla game). Changing a few towns to castles helps France.

~~

For the other factions on the vanilla map, early expansion/survival problems are mainly caused by pathing glitches.

[nikolai1962]

Nikolai, very very good investigation. It will come extremely useful I think, to everyone who reads this thread. Here I brought your attention to an aspect of the campaign map, that might be worth investigating in conjunction with all this.

Replied originally in the thread you linked but on re-reading this it looked rude not to reply directly. Apologies :)

As to the crux of your question I don't think you can aim for a general's-only movement system unless you also have a system to make sure the AI has a surplus of generals to cities, as the AI likes to have one general per city if it can. That's even if it could be made to work--dunno.

Best way to get the AI to use more general-led armies would be to have them always have more generals than cities and the player (which they do anyway but not enough to reduce the number of captain-led ones). Not sure how to achieve this simply as an AI only thing (as I don't like recruitable generals as a player--too easy.)

Having recruitable generals on a 0.1 1, or 0.2 1 sort of basis, both in towns and castles, cheap and low upkeep, might work. Player would need a house rule not to recruit them though.

Just had one thought using events, If you made the generals recruitable based on an event and then made that event happen on turn 1 for each AI faction but not the player. Not sure if that's possible.

[nikolai1962]

Recap on the strong generals thing. Haven't been testing it explicitly, just looking out for it while testing the map, but with my weakened generals (1HP and same stats as their equivalent unit e.g feudal knights for NE), then low-star generals seem to mostly have 2-3 other units in the garrison now. High star generals are still usually on their own.

[SigniferOne]

An example of point (5) in the original post about the upkeep cost of generals. Venice in the vanilla game has four start generals and also four kids in the family tree ready to come of age (compared with say England who has three and one respectively). This means they'll be paying for eight generals quite early on in the game which is roughly 4000 florins which they usually can't afford as they don't have an easy early expansion.

Obviously generals are a big advantage as well as a cost, but if you mod things like the 2HP of the generals or reduce king's purse and then wonder why Venice doesn't seem to do very well (as an AI faction) this is part of the reason.

Good observation. Another workaround is bringinging general's upkeep down to zero, so as to not penalize factions for having more 'control units'.

This brings me on to generals and autocalc. Autocalc is heavily influenced by:

1) Number of soldiers - even if they're crap
2) Melee attack rating

(why horse archer AI factions were so useless in RTW without lots of infantry or sneaky modding)

Autocalc is also very heavily influenced by multi-hitpoint units and command stars.

(One of the reasons the Egypt and Britons AI factions always did so well in Vanilla RTW was because of this--2HP as a general plus a bunch more for the bodyguard being chariots.)

Was this changed at all in M2TW, you think?

Haven't tested that last point yet. Reduced the stats and HP on the generals in my modding for the other reasons mentioned and the AI garrisons have been a bit bigger but it may be due to some other factor.

If you've been following UAI2.20 for SS5.1, they have apparently big stacked garrisons. It seems primarily an AI file issue, not so much a hard-coded bias towards generals.

Although your latest update with 1hp generals seems to add more garrisons. But I don't think generals will be made 1hp merely to suit AI fancies, I think an AI fix will be better.

(Another aspect of this potential general effect vis a vis garrisons is that the strong stats/multi hit points of the general make it a strong autocalc unit but doesn't help much with unrest because of the low numbers in the unit. Hence possibly why the AI loses remote regions to revolt so often after taking them. They leave the general behind as garrison and move out.)

Hopefully this will be fixed with improved versions of the AI file.

[nikolai1962]

Good observation. Another workaround is bringinging general's upkeep down to zero, so as to not penalize factions for having more 'control units'.

Yep. I have general's bodyguards as zero upkeep and reduced the cost of generals down to 100 in the character.txt. Meant to add that as a recomendation :)

Was this changed at all in M2TW, you think?

I think that's the problem. I think the autocalc is based on the old system where the general's command stars gave an attack bonus to each unit so the stack thinks it's stronger than it is.

If you've been following UAI2.20 for SS5.1, they have apparently big stacked garrisons. It seems primarily an AI file issue, not so much a hard-coded bias towards generals.

Although your latest update with 1hp generals seems to add more garrisons. But I don't think generals will be made 1hp merely to suit AI fancies, I think an AI fix will be better.

Yep, not saying the General effect is critical re garrisons, just a minor contributing factor. The AI settings are more important. My main reasons for weakening the generals are:
1. I think in the battles the units often make 1 vs 1 match-up decisons and i think that is part of the general's suicide charge thing from STW onwards.
2. I think strong generals are much better for the player than the AI because of suicide charges.

The garrison bit just seems to be a minor beneficial side-effect. I wouldn't claim it as a big thing though--heavily weighted to personal preference.

[Abokasee]

Out of intrest could you findout signifcant is the faction personality (as I call it, ie: sailor napoleon, balanenced smith) on a faction, and where can I get the stripped down map?

[FactionHeir]

Out of intrest could you findout signifcant is the faction personality (as I call it, ie: sailor napoleon, balanenced smith) on a faction, and where can I get the stripped down map?

AI personalities I responded to in your own thread already.
The stripped down map is available in alpaca's stripped down mod installer.

[Abokasee]

Thankyou

[Thermal]

as u can tell im not amazing at this or i wouldnt hav 2 ask but, im modding the retrofit multiplayer version of the game, but there doesnt seem to be a descr_character txt file in the folder, the nearest thing to it is a export_descr_character_traits, strange seening as every other version or mod of the game hav both files...
any ideas which descr_character file the retrofit type use? is it just the descr_character from normal medieval? it should hav 1...
i just want to add movement point up to make 4 a better AI

[CavalryCmdr]

I answered in Modding movement rates? thread.

Increasing the movement rate will only help with some pathfinding issues with the game, with some armies getting stuck. Chances are if this is the only thing you change you will not see a notable improvement in the AI, other then perhaps improved early game expansion.

[SigniferOne]

CavalryCmdr (or Nikolai): we're having a guy make a massive new map for us, and I raised to him the concern that cities should be within 1 turn from each other for the AI to make proper decisions. Is that correct, or what is the proper guideline for how far the cities can be spaced out?

[MarkaR]

Hello,
I have read this important theme and have decided to participate in discussion.
As Nikolay 1962 said: the main problem is an unwillingness of AI to develop a city and as a consequence to employ strong and expensive units. Other problems: how to force Pathfindig to work together with AI type according to mode's strategy.
AI analyzer (we name it so) a choice of the decision for behaviour of fraction consumes the following information:
1. Military balance of units (it is possible to change it)
2. AI type (each type is fixed in engine and to change it is impossible, but it is possible to jump from one type to another changing AI label)
Points 1+2 represent military force of fraction and warn AI against an attack
3. Economy (it is possible to change and create different ways) involving AI and stimulating an attack
4. Diplomacy has limited character for AI and is subject to strong development and change in the near future
5.Pathfindig depends on complexity map (it is possible to change it) and has executive character for AI after AI analyzer will compare conclusions from 1+2 (against an attack) and 3+4 (for an attack).
In our mode “Rusichi” http://www.twcenter.net/forums/showthread.php?t=187774 we have tried to change concept MTW2 (many identical fractions with the same power) and to create preconditions for manufacturing new type of AI, meeting the requirements of our gameplay mode. As the reason to it was that we had broken in a hard way the balance of points 1,2,3 in comparison with the original game. The following points are 4,5. Usually at first the universal AI is made, proceeding from the general requirements. Maybe, as the result, we will also receive universal AI, maybe not. It will depend on limits of updating of AI connected with a game engine.
We have broken down military power of units into three classes. Each class has the factor G equal to the relation of the sum of hiring and the maintenance to military force of division (the sum of military characteristics of a separate unit increased by quantity of units in division).
So the first class of units has G from 0-0,5-those are units of the early period and AI them employs standardly.
The second class of units has G from 0,5-0,8-those are already professional units and AI has no problems with them
The third class of units has G from 1,0-1,5 are unique units and AI does not employ them in general. These units are accessible only to the player.
Thus, varying factor G, it is possible to achieve from AI a fair hiring. Frankly speaking, the big changes are needed in fraction economy in order for the effective hiring of units. There is a need of making different types of economy for fractions. The type of economy is based on the speciality of fraction (speciality of agriculture, trade, religion etc.) and is connected with the result of military class unit.
Also, it is necessary to talk about generals. AI often proposes them and makes them in a family. Generals are unique units and consequently are useful in the military plan and are harmful for the economic. To adjust their number, the second class units with characteristics of murderers of generals must be created. AI actively uses them especially in auto-calc.

[Tchouk]

A major problem, as I see it, is the fact that most AI (and associate things like pathfinding) tweaks, even full AI modifications (like Lusted's), awesome though they are, are aimed at making the vanilla game-play more interesting.

The AI of the game is smarter and harder, yes, but the game remains largely identical.

What I'd like to see, and this is something that we hope to achieve later with our mod, is an AI (+ associate elements!) rewrite that balances out a new type of game-play. An AI that changes the game-play.

A lot of good or innovative ideas are fully possible to implement, and we've even implemented some of them. But while fully possible to put an idea in the game as it is, the AI doesn't react to those innovations in the way it should. For example our villages - 3x3 regions with level 1 capped towns - great idea and large game-play tweak in theory (they are hard to defend and potentially very valuable in numbers). The problem: the AI completely ignores 3x3 regions. And while this fact by itself can and does lead to its own features, it forced us to lower the value of villages to balance the game (the idea sort of fell through )

Forcing the AI to behave "the way it should" to fit new game-play is much harder and requires a lot more work, but I think it is ultimately worth it.

[CavalryCmdr]

we're having a guy make a massive new map for us, and I raised to him the concern that cities should be within 1 turn from each other for the AI to make proper decisions. Is that correct, or what is the proper guideline for how far the cities can be spaced out?

Yes, if posible, all setlements should be exactly one turn away from all adjacent settlements for optimal AI performance. Though according to nikolai1962's map modding research actual settlement locations dose not matter as much, but the "center" of a region should be 1 turns movement from the "center" of all adjacent regions. I recomend the mapper for all mods read this thread.

@Tchouk:
The problem is that new "gameplay" elements need to be made within the scope of the AI, not the other way around. This limits the modding capability as far as gameplay it's self goes as the vanilla game it's self includes gameplay elements that are not within the scope of the AI, for example the AI is incapable of recognizing the difference between castles and cities.

@MarkaR:
Your post is dificult to read, so I may be misunderstanding cirtain aspects (I assume English is not your first language)

As Nikolay 1962 said: the main problem is an unwillingness of AI to develop a city and as a consequence to employ strong and expensive units. Other problems: how to force Pathfindig to work together with AI type according to mode's strategy.

This is not entirely true, the AI will refrain from building armies if invade="invade_none" is used when not at war (descr_campaign_ai_db.xml)and will build up it's economy and military buildings, though it will still try to maintain military balance with neighboring factions.

2. AI type (each type is fixed in engine and to change it is impossible, but it is possible to jump from one type to another changing AI label)
Points 1+2 represent military force of fraction and warn AI against an attack

This I dont really understand, the AI labels themselves can be changed by modding the descr_campaign_ai_db.xml. If you are unable to open/mod xml files I suggest googling "xml marker" its a free downloadable xml editor that I find extreamly well made, it is now the program I use to open/edit all xml and txt files.

5.Pathfindig depends on complexity map (it is possible to change it) and has executive character for AI after AI

again all mod mappers should read nikolai1962's map modding research thread (link above)

We have broken down military power of units into three classes. Each class has the factor G equal to the relation of the sum of hiring and the maintenance to military force of division (the sum of military characteristics of a separate unit increased by quantity of units in division).

I have tried a similar method, keep in mind that the AI is more interested in the cost-effect of a unit rather then a unit's actual strength, also when considering cost-effect the AI is not able to consider upkeep, only initial recruitment cost, thus the upkeep of all units should be made a set percentage of the recruitment cost. Thus to get the AI to build higher level barracks etc you must make the higher tier units more cost-effective rarther then just better. Also the formula the AI uses to determine cost-effectiveness is quite a mystery to me.

Also, it is necessary to talk about generals. AI often proposes them and makes them in a family. Generals are unique units and consequently are useful in the military plan and are harmful for the economic. To adjust their number, the second class units with characteristics of murderers of generals must be created. AI actively uses them especially in auto-calc.

The AI is capable of spawning generals as needed, more on harder difficulty settings. This cannot be controled, however scripting in a refund for all generals in an AI faction (generals, not the unit associated with them) is posible and limits the impact they make on the AI's ecconomy.

[Tchouk]

@Tchouk:
The problem is that new "gameplay" elements need to be made within the scope of the AI, not the other way around. This limits the modding capability as far as gameplay it's self goes as the vanilla game it's self includes gameplay elements that are not within the scope of the AI, for example the AI is incapable of recognizing the difference between castles and cities.

Yes, but you could potentially write a script that would mimic, to a certain extent and in some cases, an AI that does recognize the difference. It obviously wouldn't be anything you could call elegant, but there are a lot of possibilities to "trick" the AI into behaving the way you want it.

[SamuraiofDoom]

By any chance does the add_money cheat screw up the game if your starting of as a small faction. as you said there are certain estimates on hoe much money you will get a turn and hoe much you can build and recruit.

[CavalryCmdr]

@Tchouk
Indeed, scripting can achieve a great deal of results, but in my experience the more scripted behavior the more open the AI is to player exploits.

@SamuraiofDoom
If your talking a player-controlled faction, that depends on how you spend the money, there are no inherent problems with it. However if your talking an AI controlled faction, they would recruit more units then they are capable of maintaining and their economy would crash as a result.