How to improve the tactical AI

[Cheetah]

The weakness of the tactical AI is an ever returning problem. What I would like to do here to get some ideas together how to improve the situation. I have been playing the TW series ever since Shogun and I am a bit afraid to say but the tactical AI seems to be the weakest in MTW2 (though I have to add that I have not played RTW SP that much). First of all one have to find the reasons, second we have to find a cure.

I suggest the following reasons: (a) variety of unit choice, (b) bad composition of AI armies, (c) wrong priorities for the tactical AI.

(a) Variety of unit choice: thinking about the weakness of the tactical AI Shogun often came to my mind, for the obvious reason that SP battles seemed to be much difficult and challenging. Thinking about the reasons I came to the conclusion that there are two major factors: (i) restricted availability of cavalry, (ii) same –and restricted- unit roster with AI.
(i) Cavalry needless to say favours the human player as the AI is unable to do the same cav tricks as the human player. In Shogun especially in the early stages armies were composed of only yari samurai and samurai archers (you had to have the master swordsmen event to get nodachi, and all types of cav units were further down on the road). I remember that as Shimazu I fought the Imagawa armies with only YS and SA, often my armies were only 4 or 5 units strong (and that was much!). I still remember that in my first battle vs. Mori my YS/SA army was soundly defeated by a nodachi heavy Mori army (and I had to restart the campaign realising that there is no way that I can beat that nodachi heavy army with only YS/SA). As opposed to this scenario high quality cavalry is highly available for the human player early on (jinettes, magyar cav, hunagrian nobles, vardariotai, kazaks, dvor, all kind of knights). There is hardly a faction with which it would be difficult to field a cav heavy army even in the early game. Needless to say most of the players are happy to live with this opportunity and soundly beat the militia/archer heavy AI armies with their choice of cavalry.
(ii) The variety of available units also favours the human player. Not just because it includes all kind of cavalry but because it is not hard to find a counter unit to the actual AI armies. Some factions have elite archers or elite infantry early on, it is not difficult to field superior armies compared to the AI even if someone would restrict himself/herself to inf heavy setups. As opposed to this in Shogun human and AI were fighting with the same units: YS and SA, there was no way to gain the upper hand in the quality of troops, and often AI factions got the elite units first (Mori: nodachi, monks, Takeda: cavalry). As a side effect, but important one, the AI could not go wrong with its army builds: it could produce only YS and SA, same as the human player. As opposed to this in MTW2 the AI can build all kind of silly armies and unfortunately too often it lives with this choice. This leads to our second point.

(b) Bad composition of AI armies. The AI often mass produces armies of cheap units even when it has all the city/castle upgrades to produce high quality units. For example in my recent Hungarian campaign I too often had to beat spearmen/ byz inf/militia garrisons in Byzantine Citadels’ that had fully upgraded barracks, stables, archery ranges. The Byzantine Citadels’ were better upgraded than mine (actually Barn was only a Fortress) yet I had higher quality troops (FK, DFK, Hungarian nobles) than the poorly manned garrisons. Also the AI will often mass produce any unit available in its cities. So, for example if a city can train militia spearmen and catapults then you can bet to meet AI armies composed of these two units without any archer or cavalry supporting them.

(c) Wrong priorities for the tactical AI: The AI often either tries to mimic aggressive cavalry tactics with the result of losing all of its cavalry leaving its infantry/archers open to be flanked/surrounded by the human player, or remains utterly passive to be shot to pieces.

Cure:

(a) Well, there is not much cure for this. CA will not decrease unit choice and it would be naïve to expect human players to stick to town militia to give a fair fight to the AI. It is a kind of an unfair situation to the AI but the unit diversity of MTW2 (which is IMO a welcomed feature) is partly responsible for the downfall of the tactical AI.
(b) The situation could be improved though if the AI would focus on producing high quality troops. More than that, the AI should focus on producing high quality all-rounder troops. For example the advantage of YS/SA system is not just that both human and AI had the same armies but that YS is a reasonable all-rounder unit useful both vs. cavs and infantry. So even in the later stages of the game when the human player had more advanced units (nodachi, naginata,etc) YS was still useful. As opposed to this militia spearmen (except perhaps the Italian ones) became obsolete very early in the game, and even some high end units such as DFK can be obsolete, especially in the hands of the AI. The AI should produce all-rounders that can be useful later in the game. Given the unit choice available to most factions IMO there is only one such early available all-rounder unit: heavy cavalry. Elite spears could be such units in theory, but given the inability of the AI to handle the shooting game and that only a few factions have truly elite spears/pikes early on (Papal states, Scotland) IMO it is a bad choice. DFK, DCK is a bad choice too given their vulnerability to cav charges. Light cav is a bad choice given that the AI is unable to handle the skirmishing game. So really only heavy cav remains. Actually it would be quite historical. Of course, given the faction, elite spearmen and elite archer could be produced to support the cavalry. DFK, DCK (swords in general) should be used only for sieges. All other troops should have a supporting / cannon fodder role. In another thread I will give examples for preferred AI army compositions but I don’t want to mess up this thread with that.
(c) The tactical AI should be treated as a beginner. When I fought Shogun MP and had allies with little experience our priorities were as follows: stick together, stick to the high ground. IMO the tactical AI should have these same priorities in that order. In MTW2 that AI too often wants to fight an aggressive cav fight which it cannot handle. Also it just simply ignores height advantage. It is happy to sit in the bottom of a valley even if surrounded on three sides with archers/ cav archers. IMO (1) the AI should keep all its troops together and engage in a very conservative orderly fashion starting from center moving to the flanks. That is, engaging in the center with its strongest infantry units then gradually covering both flanks with cavs, spears, flanker units (axemen). It is predictable, but it would be safe. Also it would keep the human player busy at both flank (as opposed to the current situation when the human player can do whatever he wants on the flank, either because the AI just ignores it, or because the AI has already lost all its cavs). (2) The AI should always seek the high ground. In MTW2 even when I start in the bottom of a valley I can manoeuvre my army into a position where I have the high ground. The AI should never allow this. I remember that in MTW the AI often changed its position when it felt its flank threatened. I never saw this in MTW2. I don’t think (though ofc I can be wrong) that it would be that difficult to program the AI to seek the high ground while keeping all its forces together (especially in the light that it was capable to do something similar in MTW). IMO such behaviour would give a much greater challenge and would create the impression of a more intelligent AI (which ofc it won’t be but still … ).

All in all IMO the weakest link in MTW2 is the tactical AI, I guess most of us would agree. With some relatively simple changes suggested above the situation could be improved. I definitely would like to see a more challenging tactical AI when the Kingdoms come out or even earlier if possible …

[TevashSzat]

IMHO, the best way to solve the ai problem is to model the ai's behavior and tactics off of people of varying skills for different difficulty, but have different styles of player that will be modeled after. For example, there could be a blitzer or a turtler and the ai could randomly pick a tactic at the beginning of every campaign.

[Foz]

I don't think getting the AI to field a lot of cavalry is a solution. As you pointed out, Cheetah, cavalry clearly gives the edge to the player since he is far more adept at maneuvering and using it. Much of this is down to the mobility of the unit, so clearly we want a less mobile unit favored. The only problem with heavy infantry like DFK is that they are somewhat disadvantaged to cavalry (note that working shields make them better than you'd expect - 21 armor goes a long way). The idea then shouldn't be to make the AI rely on heavy cav and try to match the player... but rather, I think it should be to severely limit heavy cavalry, so as to narrow the player's advantage from it, and make core infantry units like DFK viable as the basis for an army. In one move, you'll have put a leash on the human's ability to go nuts with all-heavy-cavalry armies, and allowed the AI to lean on a unit that really ought to be a viable mainstay of any army.

As for the overall implementation of this idea, I'm recommending rebalancing the recruitment pools. This is where ideal army composition comes into play. It's obvious that the AI will recruit 24/7 as much as possible, so the best way to control composition is to control availability. If you know what you want the army to look like, you can make the units available in those ratios, and I'm betting AI armies will come out looking largely like the recruitment ratios. The trick will be lowering overall availability enough to keep the AI producing units fairly regularly, but also to run out of the units it would otherwise overbuild. As such, I have a few guidelines:

1. Overall unit availability for a given settlement should be based on its number of recruitment slots. Ideally if a castle has 3 slots, then the total of all available units there should be 3.0 each turn. That would mean that exactly the right number of troops would enter the pool each turn... just not whole units in any given category. What I hope to have happen, then, is that the partial unit each unit type gets each turn translates into a different selection of 3 units each turn... and over a few turns, the correct composition should be achieved. Obviously it will never work out that ideally, but it seems a good place to start. In reality I think some breathing room is necessary - extra unit replenishment to allow the human player some choice. Just how much can be allowed without wrecking the AI composition, though, is something to toy with.

As a simple example of #1, lets say a castle can make 6 different units. If we want them in equal amounts, we'll set each to receive 0.5 unit every turn. They'll all hit 1.0 at the same time, and the AI selects 3 to build. Next turn, though, those three will only have 0.5 unit, and can't be built - the AI will have to build from the remaining three units, each at 1.5. If the AI keeps recruiting the max allowed, it would cycle turns recruiting the first 3, then the second 3, and end up with even amounts of everything.

One possible extension of this is placing lower caps on the pools. This is likely required, as the recruitment queue can't run constantly, so units could build up and allow too much choice, thus ruining the plan. The exact amounts would likely come from simple trial and error.

2. Cities may need sharper cuts than outlined above. One major problem is that the AI recruits militia in particular too heavily, so it may be useful to consider limiting army availability in cities to less than the number of slots per turn. Perhaps about half the number of slots to try to keep it under control. Obviously some factions get great army units from their cities though, so the more drastic measures are only really intended for factions where it's horrible to be recruiting tons of city troops, and castle troops should be emphasized. One additional benefit of this is that it could alleviate AI money troubles - if recruitment is more limited, the AI may find more money to invest in good buildings, and thus end up with a better economy.

3. Availability of lower-tier troops should actually go down as tech level goes up. This is something the vanilla game fails to accomplish. The way I've proposed to fix things, this would be necessary to promote recruitment of the better units that become available. In essence the idea would be to set ideal army composition for each city/castle tech level, and then pattern the recruitment allotments directly from that composition. For instance when England gets Armored Swordsmen at Fortress level, that settlement would shift from filling Billmen and Levy Spearmen faster to producing Armored Swordsmen more quickly. It may take a few turns for the adjustment to show, but in the end those lower units would quickly run out of stock and be replaced frequently by recruitable Armored Swordsmen.

So to wrap up, the essence of this is choice - If you can limit the AI's choices, you should be able to more or less determine what it recruits. Note that I haven't thought about this further than that: I don't know if this would be horrible for the player to deal with, or anything else... just that it's probably the best approach to getting the AI to recruit what you want it to. Of course the best thing would be direct control of the AI's recruitment policies, but it seems we can't have that. Anyway, feedback, as always, is appreciated.

[nikolai1962]

As mentioned, army composition has to be the first step. No point talking about flanking etc if the AI has all catapult armies. As the AI will spam units, controlling what each faction can spam to suit their particular advantages will make a lot of difference. It should be fairly easy to do now with the recruitment pools etc.

edit: giving all the units the same "walk" speed would help too so their formation didn't break up allowing cav/skirmishers/infantry to stay in formation.

[Brutal DLX]

Army composition is not that big of a problem, the problem is and has been since MTW (in Shogun it wasn't so much of a problem due to the limited unit roster) that the AI is unable to disband outdated troops. As such it always operates at the upper limit of its upkeep capabilities which it maxes out early on in the campaign with weak units simply because it can't build better ones at that time. Given the fact that the AI often does try to retreat instead of fight when facing a superior army and the fact that AI vs AI fights rarely see the losing army destroyed but rather just weakened, the AI WILL usually retrain these units, thus again preventing itself from producing any better units (well, at least in significant numbers) that might now be available.
Solution is either implementing a smart disband ability for the AI, or the player attacking and wiping out AI stacks without taking that faction's settlements. I usually do this and let the AI rebuild, which, at least in my experience, makes them produce armies with better troops (e.g. many varangian guards and elite archers as well as heavy cav for the Byzantines or lots of Papal guards and xbows for Papacy). Another solution would be to give the AI more income by gifting them provinces, however, that's good only as a temporary fix for maybe one or two factions, given the limited amount of settlements available.

[Carl]

Lots to say here.

I want to reply to several people, so I'll go in order.

Bear in mind a lot of what I have to say is colored by my experiences with both of my ProblemFixer Mods.


First, @ Cheetah, (in answer to your cure suggestions):

(a) This IS a problem, but only because of poor recruitment and tactical AI programming, and games these days HAVE to have verity if they are to sell. Thats why I tend to be dismissive to a certain degree of Puzz3D's comments about this. If they removed the verity, (as Puzz3D seems to want), then the game would not sell well and the series he so loves could easily cease to exist. Greater unit verity is something you have to accept for good or ill if you want to be able to play TW games at all.

HOWEVER their is one area of unit verity that DOES need an improvement, and that is unit duplication.

DFK, (although they DO perform worse than their stats suggest), are nearly identical to Dismounted Chivalric Knights/Armored Swordsmen/Noble Swordsmen e.t.c. The difference just isn't enough to justify any faction getting both. That REALLY hurts the AI IMHO as it leads to too many units of a similar type showing up, killing army verity.

(b) I don't agree with your assessment here as it is an assessment mired in STW thinking IMHO, (no offense BTW)and it had low verity and thus used a moderate counters system. In effect the units where largely similar to each other, the bonuses each unit got against other units that it was the counter to where small enough that whilst it might beat it's target it would only do so by a small amount. In effect the bonuses where just enough to make them important from a cost vs. power perspective, but not so great that they couldn't be overcome by intelligent use of the overall army.

With the increased unit verity this is no longer possible. The greater unit verity means a greater variance within each class type, (S & S being an example of a class type), this means that the gaps between the various class types have had to be increased creating a hard counters system.

In this system an all rounder unit simply doesn't work because it is so powerful just for it's all rounder abilities that it has to have an insane price tag. it's possible, but it becomes far too expensive to form the core of an army. What you have to do is make the AI use a balanced army and use it well.

The closest thing to an all-rounder however is S & S units as these units are able to resist light Cav, hold Medium Cav to a draw, beat any other non-pike/2-hander infantry available head on, and lastly they are very missile resistant. Their biggest issue is that Heavy Cav and 2-Handers can destroy them badly, and Medium Cav may get destroyed itself, but said medium Cav will also destroy them.

However their high defense makes them relatively resistant to mistakes and as a result they make excellent line infantry.


(c) The problem with the tactical AI is that it's programmed in a way that is appropriate to STW where units where not hard counters, and where the variance within a single class was much lower. It treats all missile units as the same in priority term, all Cav as the same, and all infantry as the same, it doesn't even differentiate between Spears, 2-handers, Pikes, and S & S infantry.

The AI needs much better target priority if it's to do ANYTHING right now as it considers border Horse the equal of Gothic Knights ATM.

Here's how I would do it.

Their would be a separate Target Priorities file, here's a notation based version.

File Header

Faction Name: ; this Section should contain the name of the faction to which the subsequent entries are relevant.

Target Faction Name: ; this line should contain the target faction to which this data is relevant, every faction should have it's own entry vs. every other faction including itself, (for use when facing rebels).

Unit Name: ; this line should contain the name of a unit that is part of the unit lineup of the Faction listed in the relevant Faction Name line of code. Each unit a faction has MUST have exactly ONE entry in each entry group.

Target Unit Class: ; which class, (S & S, 2-Hander, Pike, Heavy Cav, Medium Cav, Light Cav, Light FA, Heavy FA, Composite FA, Light HA, Heavy HA, Composite HA, Spear, Peasants, Anti-Personnel Artillery, and Anti-Building Artillery are the default classifications. New classes can be added at the top of the file).

Target Priority [S & S]: ; the target priority of the named unit against units in the named class, (S & S in this case). The units must be listed by unit name, and must be listed in descending order of priority, (i.e. the first listed entry is the unit it will look for first when attempting to match the listed unit up against the listed unit class). Each Class of unit must have it's own target Priority entry but their is no requirement to list these target Priority Entries in any particular order.

And Below an example entry:

Faction Type: England

Target Faction Type: Byzantium

Unit Name: Billmen

Target Unit Class: 2-Hander

Target Priority [S & S]: Dismounted Byzantine Lancers, Byzantine Infantry, Dismounted Latinkon, Town Militia
Target Priority [Spears]: Byzantine Spearmen, Spear Militia
Target Priority [Light FA]: Archer Militia, Peasant Archers
Target Priority [Heavy FA]: Trebizond Archers
Target Priority [Composite FA]: Byzantine Guard Archers
Target Priority [HA]: Sythikon, Byzantine Cav
Target Priority [Composite HA]: Vardariotai
Target Priority [Medium Cav]: Byzantine Lancers, Militia Cav
Target Priority [Heavy Cav]: Latinkon, Kataphraktoi
Target Priority [Anti-Personnel Artillery]: Ballista
Target Priority [Anti-Building Artillery]: Trebuchets, Catapult

The above Target Priority is just part of the mix however. this gives the main AI a good idea of what to use against what in particular faction vs. faction lineups. Which of the various categories of Target Priority should be used and 50 other factors still need to be dealt with.


The first thing about the new master battle AI file is that it will need to have separate entries for EACH of the Factions as different factions will tend towards different army compositions and fighting battles with a mostly HA force is very different to fighting them with a mostly FA force, or a mostly Spear force.

In addition each faction will need several AI sub sections itself.

You'll need separate entries for Offensive and defensive Field and Bridge battles as well as separate entries for attacking and defending sieges and sallies.

For each of these AI setups their will then need to be a section listing what order to reference the Target Priority entries of a given unit class in.

For example for English 2-handers in most situations it should try to have them go after S & S units first, Followed by Spears, followed by other 2-handers, Followed By Composite FA, followed by Heavy FA, followed by Light FA, e.t.c. This list is known as The Target Priority Order


After that part their should be a series of Formation rules. These in effect are used to cause 2 things to happen. First they make sure the units stay in formation and close to each other. Second they modify The Target Priority Order in such a way that they prevent the various units moving in a fashion that causes them to leave themselves or others vulnerable.

For example Schlitrom Capable spears in Schiltrom should be placed on the Flanks whenever possible. SO even if the unit match-ups dictate that perhaps the Billmen should go on the extreme left to counter a unit of Dismounted Byzantine Lancers out that way, they will not in fact put them further out than the spears.

Likewise it is a very BAD idea to place anything but spears/pikes in the front rank of your formation as otherwise the enemy Cav can simply charge straight through your line. The 2-Handers/S & S should however be placed in the second line, preferably with the 2-handers towards the flanks.


Likewise, the target Priority order may be altered based on what the enemy is doing and what type of battle is being fought.

For example 2-Handers should normally target S & S units first with Heavy Cav being right near the bottom of their list and they should form the second line of a formation behind the Pikes/Spears. If Fighting an offensive battle battle and you are near the enemy and their are not enemy heavy Cav close enough to react in time it would not be inadvisable to have the 2-Handers, (and S & S units), charge through the spearmen and engage their targets whilst the spear do a 180 turn and present their spear points to the rear and sides, thus guarding the flanks and rear of the ongoing infantry fight from the enemies flanking attempts.

You could include reserves in the formations too and these could be accommodated inside the spear shell where they can either reinforce the ongoing battle, or help the spears out as they become engaged. The Heavy Cav could also be programmed to come in at some point, either to attack anyone that tries to flank, to try to flank themselves, or to attack forces attempting to attack their own flanking forces.

Lastly the exact plan of attack/defense and deployment position should take account of both the existing terrain, weather, and the direction from which reinforcements, both friendly and enemy will enter.


The core of the system however is a detailed target Priority system that is then coupled with an equally detailed formations system that takes complete account of terrain effects and specific unit capabilities. Lastly, BOTH, of these elements can then be modified on the fly by the enemies actions and reactions allowing the AI to adapt.

The current AI has no adaption ability, and only very basic target priority and formation systems. As a result it totally fails to accomplish ANY of the 3 points above.


@Foz:

1. This isn't a bad point Foz, but it's my experience you can set the replenish rate up as high as 0.7 and still get decent variation. You may be able to go higher but I haven't tried bigger numbers.


2. I find that if you add + 1 to the number of recruitment slots in castles the AI concentrates much more on castle units so long as they are actually better than the city units available.


3. Totally agree and that is precisely what I've done in my re-build ProblemFixer. Give me a few minutes and I'll upload and post a link to the modified file for you.


@Brutal_DLX: your comments are only true so long as it's easy for the AI to recruit large quantities of the trash, if you follow Foz's suggestion 3 I find that you'll see massive amounts of high quality units. In my Rebuild ProblemFixer I've faced Danish Armies utilizing mostly Dismounted Chivalric Knights, Huscarls, Dismounted Huscarls, Scouts, Obudshaer's, and Norse Archers, with the odd unit of Viking Raiders/Norse Swordsmen mixed in. Yes Town Garrisons still tend to be Spear/Town Militia a lot, but that has a lot to do with the fact that these seem to take longer to update, but they eventually do update too. That and the later units are pike units which the current recruitment AI treats as the worst units in the game.

A much better Recruitment AI and Garrison AI needs to be created IMHO.

[Cheetah]

Army composition is not that big of a problem, the problem is and has been since MTW (in Shogun it wasn't so much of a problem due to the limited unit roster) that the AI is unable to disband outdated troops.

I agree that one has to make sure that low tier unit disappear as more advanced units become available. IMO the simplest solution is to deny the abality to retrain low tier units. So for example if a citadel can train CK/DCK then it should not be able to train or to retrain FK/DFK. Eventually the existing low tier units will be used up and thus will disappear.

[Cheetah]

I don't think getting the AI to field a lot of cavalry is a solution. As you pointed out, Cheetah, cavalry clearly gives the edge to the player since he is far more adept at maneuvering and using it. Much of this is down to the mobility of the unit, so clearly we want a less mobile unit favored. The only problem with heavy infantry like DFK is that they are somewhat disadvantaged to cavalry (note that working shields make them better than you'd expect - 21 armor goes a long way). The idea then shouldn't be to make the AI rely on heavy cav and try to match the player... but rather, I think it should be to severely limit heavy cavalry, so as to narrow the player's advantage from it, and make core infantry units like DFK viable as the basis for an army. In one move, you'll have put a leash on the human's ability to go nuts with all-heavy-cavalry armies, and allowed the AI to lean on a unit that really ought to be a viable mainstay of any army.

I think we agree here. I just think that in general there are two ways to deny the cav superiority from the human player: (a) make the AI to build lots of cavs (like mongol style stacks), (b) or limit the human player. I agree that the later option perhaps preferable to the former.

ideas about recruitment pools ...

I also agree that something have to be done about recruitment pools, though I have a bit more radical idea, namely to use castle recruitment pools as a max limit for that kind of unit, like we have max limit for priests, spies etc. More about this later.

[Didz]

The AI should monitor the players game play style and adopt the appriate counter tactic's to overcome them.

[Jambo]

One of the things I always thought the AI should be doing is building units in cities where there's an immediate an obvious threat to the city; like an approaching enemy army or a build up of units at the border. As such this is the AI reacting to a situation.

The number of times I've waltzed a large army into AI territory only to be confronted by a rather defenceless city containing only one AI general are too many to name.

[Cheetah]

IMHO, the best way to solve the ai problem is to model the ai's behavior and tactics off of people of varying skills for different difficulty, but have different styles of player that will be modeled after. For example, there could be a blitzer or a turtler and the ai could randomly pick a tactic at the beginning of every campaign.

IMO some styles fits the AI better than others. For example good blitz (rushing) is not that easy. For example, ATM the AI looks more like a noob rusher than an expert one (sending forth and wasting all of its cavalry well before its infantry could get there). I would say that if that is the AI "knows" about rushing then let it stick to a more conservative style.
All in all considering the limitations of the AI some styles may not be suitable for the AI at all.