Game AI stuff (1.5)

[nikolai1962]

Ty

~~~

One of the other places stalls happen semi-often because of how the simple map is set up is shown here.



A stack seems to head off from Rome with Damme as its target while Ariminum is under siege. The stack sieging Ariminum assaults, loses and retreats but it fine. However the moving stack changes its orders and sieges the now open Ariminum. *If* that stack then later assaults and loses it will retreat to the spot shown in the image. If no-one else sieges Ariminum i think it just goes back (doesn't happen every time so hard to be sure) but if another stack takes over the siege then the original Damme stack goes back to trying to get to Damme from the retreated spot but stalls.

The second image (despite my poor graphical skills) is an attempt to show my theory of why it might stall here. If the game tries to path find by comparing the distance from target from the tile the stack is on to the distance from one of the tiles the stack could move to, and *if* the path is sort of sideways like it is here then the distance will only be reduced a tiny amount. Maybe there is some kind of rounding error that makes the distances equal so the stack can't find a tile to move to that reduces the distance?

Path-finding will be hard-coded so doesn't really matter much why it happens but anyway :)



So anytime an ai stack has a target settlement but breaks off on the way to fight a battle or take over a siege it then has to find a new path to its original target. This is probably what i noticed before while watching the vanilla map. Only occcasionally would an ai stack seem to stall after fighting a battle. The ones that were on rebel/enemy chasing duty are generally fine but the ones that have a specific target can occasionally be messed up. On a normal map there may be other things happen that clear the stall but i'm sure in some situations it causes the faction to stall as it does with this map. Damme is targeted but unreachable by the stack ordered to take it so the ai doesn't send another stack.

Third image was meant to be something else but just shows my current attempt to get past this by making the blocking terrain more rounded with woods instead of forest at the corners.



I think i saw this happen on the vanilla map when ai stacks lost an assault against halicarnasus and retreated east but hard to be sure as so many thinhs happening at once on a normal map. Now i know what to look for when i go back to it maybe it will be easier.

This "retreat" stall (happens mostly with long retreats) is the biggest cause of the ai getting stuck once the general path-finding problems between settlements are fixed.

The preceding image was meant to show the other semi-regular type of stall that happens. I have no real idea what causes this one. Here a stack moves out of a city to the "sentry" position. Usually the stack then either moves off next turn or waits while it gets more units from the city but in some cases they just stall. If you use move_character the stacks goes off to a target settlement (didn't work here as there were two Captain Aulus) which seems to imply that stack was given a target settlement to attack and it is a path-finding bug again but it happens when there is almost a straight line to the target (Damme again in this case).

Maybe it is waiting for a unit from the city but the city is too far down the queue for unit-building so it never gets built? Dunno.

There's a 3rd type that happens too inconsistently to have much clue what causes it. In that case the stack doesn't seem to have a target settlement it just stalls. move_character doesn't unstick armies like this. Only current theory is that it is somehow caused by man-of the-hour events as the ai gets a ton of those with all its captain-led armies? These don't stall the faction though as the stacks don't seem to have a target settlement.

Next step is to add more rebel garrisons so the ai loses more sieges so i can see a consistent way to fix the "retreat" stall and then go back to the other two afterwards.

~~~

Brigands

Image shows what happens on first turn when you have brigand spawn rate set to zero :)



Ai never expands more than one region with that many brigands. Obviously this is an extreme case though.

As the brigands are so random you'd need to test a lot of times to be sure and i only did it three times with brigands set to 28 and 99. The roman ai fills the land part of the map a bit faster with brigands on 99 but not as much as i thought. That makes unrest move up to the 2nd spot after path-finding as the biggest AI handicap (imo).

Playing on other maps I got the feeling there were more brigands with the same setting but higher difficulty levels but watching this smaller, simpler map that doesn't seem to be true so the extra cash the ai gets on hard and very hard makes it fill the map slightly faster too. (i'm guessing that is the reason.) So i think hard and very hard counter-balances the brigand handicap (on this map).

As the ai consistently only uses one ship to transport their armies, and pirate fleets can be so strong, the speed of getting the islands varies massively. I gave up on that and only tested the land expansion speed. AI naval expansion needs less or weaker pirates imo because the AI doesn't build proper fleets often enough.

~~~

I think CA needs to move to hexagons :)

[nikolai1962]

Monkwarrior,

I started looking at the carth-roman thing on the vanilla map before i got side-tracked onto my simple one. There were a bunch of things i thought needed to be done to the map to make it work right (from the carth side) which i'll point out here. Your map is probably different but it may have some of the same problems. I only changed the first one (sahara) as i didn't know how to edit the other map files then.

1) The AI prioritizes adjacent land regions. For carth this means they spend a lot of time in the desert both south and east and especially if they get into a war with egypt, which they nearly always do. This is a big waste of time imo and stops carthage ever seeming like the regional superpower it is supposed to be. My solution as shown in the image was to make a huge sahara that touched the sea so it was the border region with egypt. Siwa and nepte were too far away from each other for egypt and carthage to ever fight whereas nepte was close to the coastal regions so carthage could take it without a long trak into the desert. I also made nepte low population/fertility and made desert regions (hidden-resource) unable to build lots of pop growth stuff like farm upgrades. This meant it wouldn't keep rebelling, as that was another distraction for the carthage ai.



With the carthage ai undistracted to the east and south all it needs to do to become land-locked and therefore naval-orientated is to finish expanding west along the coast.


On the vanilla map it seemed the cartage ai sometimes (not always) had a problem getting to Tingi. The bends in the road shown in the second image are the reason (i think).



I added a region between Tingi and Cirta because at first i thought it was a distance problem but it still happened some of the time so now i think it is a path-finding glitch related to bends. You need Carthage to take Tingi as Tingi-Corduba works as a naval expansion path to Iberia. I know now that moving Tingi, that ford to the east or something else should clear the path-finding glitch and let carthage target Tingi properly every game without stalling

With Tingi taken, and the huge sahara blocking expansion eastwards, Carthage becomes a naval expansion faction (ignore sicily for a minute). Its closest naval expansion target is Iberia and so once these things are done Carthage should always fight hard to expand in iberia.

There is another map glitch here though. When carthage has been pushed out of iberia and is re-invading from Tingi they sometimes land near where i placed the carthage army in the picture. These stacks sometimes seemed to get stalled. After my tinkering with the simple map I'd bet it is because of those mountains between the landing zone and corduba.

So I'd say to make carthage act like carthage on any map you'd need:

1) A huge sahara (or something) blocking egypt and carthage
2) Smaller/closer desert regions to let carthage take them without long slow journeys.
3) Make sure those inland desert regions don't rebel too much by fixing the pop growth.
4) Make sure there is clear path-finding along the coast.
5) Make sure there is clear path-finding between the African and Iberian regions that are closest together.

Lastly i saw a glitch on Siclily once where carthage had lilybeum and syracuse and got stalled moving to messana. I think it is the bendy road shown in picture below.



When i get back to tinkering with the vanilla map i think i'll make sicily more triangle shaped with the mountains in the centre so there are three straight-ish roads between lillybeum-syracuse, lilybeum-messana and messana-syracuse. I think this will make sure the ai doesn't glitch here.

Apart from making messana-italy have a clear naval expansion path (or a land-bridge) i think fixing these things (or the equivalent on other maps) will make the carthage ai much better.

I think it will be much trickier to get the roman ai acting like historical romans though.

[Duke John]

Good stuff!

Months ago I also looked at the strategic AI and particularly at how it didn't merge its armies. I found that a small amount of movement points was bad as it meant that an army could not reach its target friendly army. Increasing the movement points greatly reduced the amount of armies as the AI armies could reach each other within a single turn. I would advise to give AI character a standard increase in MP by the use traits.

There is also the problem of armies not merging when the total of 2 armies is greater than 20. So complete stacks are rare.

All of this was tested on 1.2 so things might have changed.

nikolai,
have you tested wether lots of movement points (MP) decreases pathfinding problems? If so you might want to test giving AI characters a movement bonus trait when they have not used any MP for a few turns and are not in a settlement; they are stuck. The sudden increase in MP might open a different path.

[Monkwarrior]

Hi again.

Fortunately we are working in complementary aspects of the AI, so I think we will advance faster sharing our experiences.

This post is to illustrate the findings reported before.

1- Behavior of a superfaction (different from the senate)

In this experiment, julii were superfaction of other factions, such as greek_cities and this is the result.


I had never seen so large concentration of armies, and julii were unable to attack the gauls on the north. Their only expansion was Liguria, that started as rebel region.

In the next experiments we changed the system, and dacia (cisalpine gauls) were the superfaction of julii. In this way the romans have no room for expansion by land, and they are more prone to sail.

2- The need for a strong motivation

As I said before, the presence of a roman fort with a full stack army was completely useless to attract the romans to Iberia.
But we tried to make the same giving the romans a single settlement in Iberia.
In that case, romans are able to send troops by sea.


Then, the romans begin a war against the surrounding faction (illergetes), which is at the same time the creator and the culture tag of that region.
This is complemented with expeditions by land, as some roman army was able to cross the gauls territory, including the Alps, without attacking or being attacked, just to arrive to Iberia to fight against the illergetes.

3- The expansion

Once that faction has disappeared, romans turn south, to the zone controlled by carthaginians (probably due to path finding questions), and they meet in Saguntum (as it happened historically).


However, the expansion of turdetani (white) was too far, in comparison with the historical behavior. They are included as allies of the carthaginians, and limit their expansion to the coast (which is also correct).

In the last experiment (for the moment), the turdetani were considered as the superfactions of carthaginians. The starting position served as shield against lusitanians, and at the same time drove the carthaginian expansion to the coast.
The result was nearly the same, but with the limitation in the turdetani expansion.


In some days, I will post the new experiments with superfactions, alliances and "motivations".

@Duke John
In those experiments MP are 95, but they are increased by large amounts for admirals, spies, diplomats and generals through traits.
I knew some days ago that the limit of 100 is no more working in RTW 1.5, and I will try with higher MP also for captains, just to see if there is any difference.

Cheers.

[Coldfish]

Impresive Work ! nikolai1962

I think you need an award for all your help given here on these forums

Can you upload that mod please ? I like it alot !! especially the removing of the britons from the mainland...and Sahara as a single region

Again.NICE WORK

p.s. What are the Thracians doing in the Alps mate ? But I'm guessing that Is just a temporary "training" map

[nikolai1962]

Impresive Work ! nikolai1962

I think you need an award for all your help given here on these forums

Can you upload that mod please ? I like it alot !! especially the removing of the britons from the mainland...and Sahara as a single region

Again.NICE WORK

p.s. What are the Thracians doing in the Alps mate ? But I'm guessing that Is just a temporary "training" map

Ty. The map files are in a permanent state of experimenting at the moment but if i ever get all the pathing right i'll probably upload for sure :)

Lol, the thracians are being the arverni as i was originally trying out variations for extending the government idea in EB so i'd started to make the factions match the EB ones. As i was testing with FOW off instead of playing i started to notice all this weird AI stuff.

Deep breath, long post coming up.

[Coldfish]

I see..well thanks! I can't wait

[nikolai1962]

After getting the idea of a stripped down strat file to test path-finding i had to try it out on the vanilla map.

Made a provincial campaign.
Removed senate, scipii, brutii and numidia (as i was going to look at carthage path-finding first).
Gave each faction their original capital as their only settlement, same size, buildings etc.
Starting leader with thier cavalry as only army.
Brigands and pirates set to 99.
10 happiness bonus on governor buildings to stop distractions caused by rebellions.
Resources, landmarks, diplomatic attitudes etc all gone.

Might be a better way to set this up to test the path-finding but the above was what i tried this time.

Only changes i made to the base map at first were extending sahara to the west and moving the briton port as i'd sort of tested them before.

The result was this.



Running it for just a few turns you start to see some places that probably have problems straight away.



The AI likes to go for the closest adjacent land target (modified by garrison, owner etc) which should have meant Chersonesus, Bordesholm and one of the two near the armenian capital would have gone pretty quick but they didn't. This was a consistent thing when i ran the first few turns a few times. The armenians always went south to hatra. The germans never found their way to bordesholm from south or west, only from the east. Chersonesus is generally messed up from all directions.



I'll go back to them another time as this was mostly about carthage.

The carthage AI starts fine, build a small army, sends it to nearest target (Thapsus), takes it. That army leaves a garrison behind and heads for next nearest (not neccessarily nearest to thapsus). Depending on how i'd changed the map over the course of this testing it either went to lepcis or Cirta. Meanwhile carthage has build a second army that goes to whichever of lepcis or cirta hasn't been targeted by the first. After those first 4 are taken it all goes wrong.

The pathing from carthage to dimmidi and nepte goes a very weird way round with stacks getting lost on the way.



(nb those two carth stacks had been stuck there since near the beginning.)

All the time it takes is time the iberians have (in the normal game) to kick carthage out of iberia. Some stacks seem to get lost during this journey too and end up wandering around lepcis permanently. They get constantly re-inforced from lepcis but do nothing. Ai stacks are nearly always doing something unless they or the faction is stuck in some way. If they're not moving or if they're moving back and forth in a repeating loop then they are usually messed up. Not always but if you watch it enough you learn to see when it isn't so.

When the small starting stacks finally reach dimmidi they assault and lose a lot and bounce off. Most of them seem to not be able to find their path back anywhere so when dimmidi is finally taken they sit around dimmidi getting re-inforced from there and the stacks will get bigger but don't do anything. They'll attack nearby rebels but won't chase them if they are outside the dimmidi region.



This shows the same thing but earlier before the stacks got huge but i drew pretty lines on it so i uploaded it anyway.




Not all of this will apply to the normal game, though the pathing to dimmidi and nepte will be, in the normal setup the ai may send bigger initial stacks and so end up with less lost ones. Really meant as an illustration of how the path-finding can be so critical to the ai, especially compared to a human player who has zero path-finding problems unless they're playing with their eyes shut.

This was all too painful to watch and also i found out the distance from siwa to nepte didn't stop the egyptians targeting it when it was still rebel, only if the carthies had it. Which they didn't as they kept getting lost.

So time to mess with the map :)

(I wasnt going for realism or aesthetics here just trying to see if simple map changes would make the ai perform to its best (which i actually think could be pretty scary with path-finding fixed)

So i lost gaetulia/dimmidi and extended the sahara over their old terriotory and moved nepte close to cirta hoping the ai would target it from there and not go the long way round. Also moved cirta/thapsus a bit to smooth out bendy roads, moved a ford etc, added some hills to push roads in certain directions to make them more rounded with less sharp bends etc. All of this is trial and error and often makes things worse but eventually it works.





Once this worked the land expansion went smoothly. A human player would still do it faster but not by as big a margin (imo)

As you can see on the minimap of one of those they then landed on sardinia.

Note. It seems that the ai doesn't build ships unless one of its regions can see a naval expansion target. It also doesn't seem to build them *until* it needs a new expansion target. None of the ai factions who weren't land-blocked (sort of) built a ship apart from the iberians who built one and landed a small army by tingi during a brief cease-fire. The ceasefire broke soon after and they never built another. They may not actually land an army but they never built a ship unless they had a target for a moment when they were temporarily blocked.

Carthage can *see* sardinia and built a ship and off they went. One of the images also shows one of the times the ai will have a lot of stacks hanging around. I thought it was because they were stuck because of either path-finding round that mountain (as the ai will move stacks to the capital when they have nothing to do) or the carthage ai couldn't see another naval expansion target and were therefore stalled.



What it actually was was them waiting for a ship. An odd place to wait admittedly (probably because of my mountains, might be better to have a path through them to thapsus). The holdup was the ****ing pirate fleets. Carthage would build a ship every turn and move it to its "waiting for an army" spot and they'd be constantly wiped out or retreated by the swarms of pirates. Again, the pirates wouldn't have been such a big problem if all the factions had their starting fleets but it still shows how important they are at messing up the ai naval expansion. This stage of the testing got too annoying as 10-20 turns went by before carthage managed to get a ship off to next target. The pirates get too strong even set at 99 because the ai uses small fleets and loses a lot.

I went and changed the EDU so carth ships started out as exp 9 but even then it took a while before they cleared a space (mainly because carthage got big enough to build the bigger ships i think). AI naval expansion needs weaker pirates badly imo. Setting the chance to 100 in the rebel_factions file seemed to help a bit (deleting the pirates section causes a ctd).

Anyway all the changes gradually made the pirates less of a barrier and i could see where carthage would go next.

Carthage was the only place yet with a port so it would only be places Carthage could *see*. (A large part of this was to see which regions could be seen as a naval expansion target from different places.)



So Carthage can see Lilybeum and Caralis and sees Caralis as closest (i am assuming that is why they went there first as caralis had a full rebel stack).
They only go by see when they are blocked on land in some way but at least the routes are there. Oh yes, i put a fordless river between egypt and carthage for this just in case the far north-western nepte wasn't enough to keep the peace with this odd strat file setup. I was only interested in what path-finding glitchs would make carthage the pushover it usually is.

Once on sicily they have adjacent land regions so that is it for naval expansion until sicily is taken (or in a normal game if they were at peace with another faction on sicily). The path to syracuse was no problem.



(notice on minimap the regions the ai factions are having problems finding. no-one ever finds tribus sakae it seems. I wasn't watching those areas as i clicked end turn but if those regions are targeted but the attacking stacks are lost/stuck the whole faction might be stalled.)

After this is all went wrong again.



The army that took syracuse attacked and lost the battle for messana. Syracuse-messana can see each other and lilybeum-syracuse can see each other (when i say see i mean find the path) but lilybeum-messana can't, so once the first army retreated that was it for the carthage ai. If this was the normal vanilla game and carthage had been pushed out of sicily and spain because it was lost in the desert, and then made a come-back this is where the faction would stall again.

This was painful to watch too but i did for a bit and it was quite funny in the end. The carthage ai kept sending re-inforcements by ship. The re-inforcements were just as stuck as the others. They landed more stacks and began to fill up all the available coast tiles. Bit by bit they got closer and closer to messana as new stacks were landed and bumped previous ones up the road. Finally they hit the jackpot and a bumped stack got round a bend or something and sieged messana. (going by the date on the images i think that took the AI about 40 turns.)



Made me laugh anyway.

Short-lived though as they lost and retreated back to syracuse and stalled too.

Pointless image but i added it anyway. I tried to move a bunch of them next to messana but only one of those stacks (maybe more) is the prime attacker. The rest are supporters and back up if attack fails so i needed to guess which was the right one. Gave up after three.



Sicily needs to be fixed for the ai to work. I think given the shape any small scale sicily needs to be triangular with the cities at the apexs(?) and straight-ish roads between the cities. I slapped a lot of mountains down in the middle and made the edges of the islands have at least two pixels of clear land for the road to go down.

Not very pretty but would it work?



No was the short answer. Needs to be tinkered with more but better to do that with a strat file set up with carth (or another faction) just on sicily so you can change the map and test it straight away. My current one sort of works in that carth now sees messana from lilybeum but i messed up their path to syracuse. But as messana sees syracuse as i hadn't changed that side of the map and as the carth armies are coming from lilybeum they don't get stuck as they go to syracuse via messana. So that was good enough for now to carry on testing the carth stuff.



None of the cities on sicily can see italy and vice versa (though i intend to try and make messana and croton see each other eventually) so no more landings here. If there was a land-bridge carthage would now start going into italy (i prefer the naval option though when it can be made possible). So next stop palma.



I wasn't paying attention here and didn't see where the ship was built that took this army to palma (that is the region that can *see* it). I *think* it was caralis as there was a port there now and i've seen the vanilla julii jump to palma as long as the senate had already got them to sardinia. Will check sometime by starting carthage on sardinia.

Then it stalled for a bit. Not sure if that means they had no more naval expansion targets from theor current position or if they iberia was already picked and they were waiting for tingi to finish its port. Either way an army moved up to tingi and waited. A few turns later the port was done. Next turn a ship appeared. Then off to spain.



With no desert distractions the ai was doing a good job focusing its forces.



Then it all went wrong. Not at all surprising now i look at iberia after tinkering with the path-finding. I think it happened when a carth stack lost a siege against numantia and retreated behind a mountain. There were too many roman and carth stacks moving and fighting to spot the glitch but whatever it was the faction stalled and it stopped sending armies to iberia.



(This is an exception to the rule about capitals. The ai only congregates around its capital when it has no targets.)

I gave up here as testing the path-finding in iberia is going to take a while and it will be better done by starting carthage in corduba without any other factions so it is easier to seperate the general map glitchs from battle glitchs.

As an indication of how messed up the ai can get in iberia:



All but one of those (8 or 9) roman stacks are stuck (including the garrison in scallabis). The romans can't find the way to the last carth city and the carth garrison there is stuck too. The whole faction wasn't stuck because it had got big enough to have multiple targets which included ones in the east of their empire. A quick scan of the map showed that was about half the roman army.


Couple of other things i saw before closing down.






~~~

Anyway i think this all shows.

1) The AI factions are massively handicapped by path-finding problems which can be found and fixed.
2) Naval invasions work fine but only in certain specific directions which need to be found by testing (and only when the faction is land-blocked physically or by alliances).
3) This is quite a good way of testing the path-finding (though probably much improvable).
4) I like linking images a bit too much :)

[nikolai1962]

Monkwarrior,

Using superfactions is definitely something i wouldn't have thought of. Seems likely it could work well in the right circumstances.

The AI always seems to prioritize adjacent land regions so it is hard to make it go to sea. One thing i tried once was make a saguntum equivalent town in iberia have romans as the faction creator so it sometimes rebelled to the romans. I want them to sail there though :)

The other thing you could try to control expansion is tweak the regions to make sure the closest one is the one is in the direction you want the AI to move or have a starting army on the border of where you want them to move. Once they are at war with someone they will follow that route mostly so if there was historically a coastal tribe and you made them closest to the carthies then that might work out. Doesn't help with the romans though.


edit: looking at your images again i'd say the first one would work best if the map was tweaked so rome could path-find to sardinia/corsica then on to palma then spain. I'm sure that is possible after my experiments. Big problem (apart from finding which tweak worked) would be making it so they saw palma but not carthage itself as carthage is closer.


I couldn't resist trying out the idea of a stripped down strat file on the vanilla map to test path-finding for carthage. I'll post the results once I've uploaded all the images. Some of it may be useful.

[nikolai1962]

Good stuff!

Months ago I also looked at the strategic AI and particularly at how it didn't merge its armies. I found that a small amount of movement points was bad as it meant that an army could not reach its target friendly army. Increasing the movement points greatly reduced the amount of armies as the AI armies could reach each other within a single turn. I would advise to give AI character a standard increase in MP by the use traits.

There is also the problem of armies not merging when the total of 2 armies is greater than 20. So complete stacks are rare.

All of this was tested on 1.2 so things might have changed.

nikolai,
have you tested wether lots of movement points (MP) decreases pathfinding problems? If so you might want to test giving AI characters a movement bonus trait when they have not used any MP for a few turns and are not in a settlement; they are stuck. The sudden increase in MP might open a different path.

That sounds like a good idea :)