Greetings Ladies & Sires,
I’ve observed the behaviour of the AI after the Sultans Ascend major update and took enough notes to be able to write about it. Furthermore, I feel the urgency of doing this after witnessing its current status.
It’s not good, sadly… 
This in the perspective of a 2 years release. There have been updates, important ones as well, but clearly there still are big issues in this department and I don’t think there’s much attention to it. The AI isn’t useless, it can be a challenge for less experienced player, but we’re far from that “deep learning definitive-challenge AI” that was anticipated… It’s fun but you can’t look too closely to it.
Beware, it’s a long one! 
But first a few requests to the Devs.
Unambiguous Naming
Give more identifiable names to the AI players, at least something on the lines of “civ_name # A.I. difficulty”, exempli gratia “Chinese 2 A.I. Outrageous”, to keep the closest to the current formula. Maybe “Outrageous Chinese A.I. 2” sounds better. Still a mouthful.
Choosing randomly from 8 (at least) historical leader names for every Civilization would be ideal, followed by the difficulty in brackets, like “Genghis Khan (Hardest)”. This would also be in line with AoE tradition, so why not?
(Please add little Civ flags near player names.) - Thank you! 
Tributes
Teach the AI to request and send Tributes to AI and human players alike.
Surrender
Have the AI drop the ball instead of dragging the game on. Provide at least some base scenarios like:
- Surrender when, in team games, you’re the last “alive” and there’s too big a difference in the teams scores.
Being alive means having an economy and some production going on. - Surrender when, not in team games, you loose all Villagers, TCs and don’t have a significant army for one last desperate attack. If the attack fails, surrender. Especially for matches without Landmarks Victory.
Commands and Communication with other players
AI commands and other interaction with human players. With the right human, this could simplify the decision making. I noticed the AI reacts to the pings but that’s very basic.
I will report and discuss the issues I’ve found, not excluding the possibility of others I haven’t noticed. I’ve observed mainly the Hardest AI, some of the cheating ones too at first.
About the testing method
There’s currently no way to have the AI play in your stead, like it is in AoE2, so my presence as a player was mandatory. I used 2 work-arounds:
- Side with one team, hide a unit in every allied capital TC and one in up to a couple Outpost in safe positions in corner of the map, or around the border. Then, since the AI no longer repair allied buildings, I torched my own TC with the this is fine cheat code (make it quick seemed to cause crashes). This minimised my presence.
- Don’t side with any team, delete all starting units and use the anti-poke cheat code. The AI almost always ignored me, acknowledging I wasn’t a threat (had an insignificant score). This prevented a Landmark Victory but I could always check the replays to see when a player had actually won, if relevant.
In any case, after that I used another cup and let the game go on its own while I did other stuff. After the matches ended (or I terminated them because they were way too long) I analysed the replays, looking for interesting stuff.
1v1s
In these games the AI is actually decent but not for its own merit. 1v1s are the simplest scenario, no team mates and only one opponent, also games tend to be much shorter. Hence, some of the AI’s deficiencies are mostly negligible.
The AI follows the same pattern: quickly ages up to Feudal, starts training troops and attacks when has enough while at the same time starts expanding with TCs to some degree, then ages up, and it sorts of repeats. This obviously if let be. Strategies like Fast Castle or Imperial don’t happen unless there are high starting resources.
That said the AI is good enough to keep the enemy under pressure as it will always prioritise training troops and attacking. This combined with a big economy, since it aims for at least 100 Villagers, is what makes it dangerous, despite the recklessness and inefficiencies.
Team Games
Sometimes last way too long, even 2v2. I saw even 10 hours games, along with several longer than 5. At best team games last between 1-2 hours.
The AI is much better at defending its allies than coordinating with them for an overwhelming attack, although it’s worse than before at recovering. Fights attract the AI players but that’s not enough to make serious coordination. Usually a team mate starts the attack and then others comes to help after the notification pops up, there’s no synchronised flanking or similar tactics, when attacks happen together it appears as just a coincidence.
Beside aiding allies under attack and protecting Wonders, the AI doesn’t seem to cooperate much. Fortification is individual (although sometimes I saw AI players put some in allied bases), I have no proof of exchanging resources (but could argue that doesn’t happen) and again I can’t see a pattern in the attacks other than following the “Ally under attack” notifications.
There has to be a whole other layer of cooperation for team games to really be so for the AI.
FFAs
With 3 players usually the duration is decent, already with 4 it easily gets hours long. The issue is that AI players don’t know how to deal finishing blows for Landmarks Victory, can’t turtle effectively for a Wonder Victory or get hooked on quarrelling over the same 1 or 2 Sacred Sites. I noticed it did a lot of contesting a Sacred Site then moving to attack an enemy base, rarely accomplishing anything.
The AI should learn that the FFA world is very dangerous! Spread and protect the Landmarks, expand towards the border of the map and change its strategy. Many preset matches feature FFAs but there’s no dedicated logic.
Water Maps
As a general rule, the AI can’t play Water Maps effectively. This deserves to be its own matter.
Most of the times the first player or team to conquer all the Sacred Sites wins, possibly after a bit of contest. Wonder Victory can also happen. Landmark Victory is almost impossible, in an Archipelago 2v2 with Landmarks Victory the match endured 20 hours before I stopped it. The AI was completely stuck.
The reason Sacred Sites are by far the most common Victory is because the AI is, as a rule, bad at transporting units across the water. Since ships can’t contest the Sites (maybe they can if Sites are on the coast but still I highly doubt the AI could do it) the AI can’t react properly and loses. It usually has to face defences around the Sites as well.
In the late game, the AI gets extremely bugged. When reaching the Imperial Age, it basically stops moving troops, no more Transport Ships, and focuses solely on Warships confrontation, eventually stopping even that and keeping its fleet in dock, idle. Sometimes an AI player loses all of its Docks and doesn’t rebuild others again, effectively giving up the water (with wood still present, let me point out, so it’s not a lack of resources). Land armies stand idle near the coast, same as the fleet. With these premises it’s clear why a Landmarks Victory gets impossible at some point and why a Wonder Victory is usually a safe bet. Only a Wonder inconveniently placed near the coast would be endangered.
The AI needs to be much better at transporting units with ships, defending its water economy, invest much more on Docks and coastal defences, create useful colonies and to not eventually bug out.
A Broken Hybrid Map: Thickets
Probably due to some bug, the AI goes crazy on that map. The AI kind of treats it like a water map, fortifying the coast and establishing awful (3 Gold/Wood) “sea” trade rather than a much more profitable land one. It’s very active in the lake in the early and mid game, getting bugged with passive land armies and Warship battles in the late. Can’t pursue Sacred Victory because apparently can’t detect the central island. If the game doesn’t end in the mid term (usually doesn’t) the only way to conclude it is with a Wonder. Land armies moving to destroy the enemy Wonder usually get melted by the coastal defences on the lake they move along of.
Multitasking
By an AI defined as Hardest, I expect top-notch multitasking, even superhuman, and that doesn’t happen. In many occasions the AI seemed very distracted. This is also an abstract way of saying that there are certain things the AI could do but doesn’t, for reasons that go even beyond APMs and priorities.
Slow Reactions
So slow sometimes! Reaction to incoming damage to moving armies, burning buildings… I cover these things later as they’ve been matter on changes in the last patch.
Fishing Boats
Unresponsive to endangered Fishing Boats, these will often just stand there if there isn’t space in a nearby Dock. I’ve seen AI Fishing Boats ignoring completely an enemy Dock defensive fire, tried fishing right in its range and sank.
Relics
It may happen that the AI ignores free Relics entirely, focusing on aggression instead. It shouldn’t have problems in tasking even just a single Monk in collecting them all, especially if it has more than one and the space to fit all the Relics.
Doesn’t retrieve the Relics from eliminated players, an easy task. The victor could make sure all Relics are released once the opponent is eliminated.
Doesn’t reassign Relics in possession, relevant for the Chinese (Pagodas), HRE (Defences & Docks) and Byzantine (Grand Winery).
Victory conditions
The AI doesn’t seem to care much for those, I had a Wonder FFA where it obstinately refused to build one, going for Annihilation instead. Not focused enough on the win. Has a natural tendency to the Landmark Victory, but not too much, since it often fails at sniping Landmarks.
Sacred Sites
Awful at contesting and capturing Sacred Sites, usually it sends a big army to a Site, even if undefended, and then moves towards an enemy base, not always bothering to take the Site first. It wouldn’t take much to leave a Monk behind for a minute or sending one in, though the AI doesn’t do it readily. It fails at considering the right objective, or at the very least isn’t flexible enough with its unit management to do it effectively. Also, It wouldn’t take much to have just a Scout contest a undefended Sacred Site. Another silly thing is how it focuses on 1 or 2 Sites leaving the other(s) be, even if extremely safe. Just send a Monk and take it already! It’s free Gold if anything…
It appears to be less prone to fortifying the Sites than previous versions, but usually does to a certain degree. Unfortunately sometimes builds too far and on the wrong side, not covering the Site itself.
It ignores Sacred Sites entirely when they’re not a Victory Condition, loosing the chance of a passive Gold income. This is less impactful since the Sites Gold generation was nerfed but this behaviour has always been present, leaving Sacred Sites for the human players taking.
Wonders
Can build a second one right after the first is destroyed (and keep trying), or not build any at all for a whole hours long game.
Random and Usually Bad Placement
Doesn’t know how to properly place the Wonder, ignores already built defences. Saw it placing one on the front of the base, although kind of near the border of the map, instead of a way better place amongst 4 allied Keeps. A particularly bad example of Wonder placement was when I witnessed an AI player placing its Wonder over its right ridge on Turtle Ridge. This also made me suppose that the AI uses distance in a beeline, not pathfinding, when placing a Wonder (and probably any building). The enemy destroyed the Wonder before its owner forces could reach it.
Defending the Wonder
In the past, the AI used to fall in the trap of only protecting the Wonder, leaving its base open to Landmark sniping. As of now, that doesn’t happen anymore and the AI is even capable of keeping an offensive going. Maybe too much though. Undefended Wonders are common enough to lead to a big waste of resources, when combined with horrible placement (front of base, slightly flank, why not…?).
To rub it in, defence works are rarely on point. Sometimes, there’s not a single Outpost around; other times, you get a couple of Keeps but without guarantee of ideal placement though, at times they get built those few tiles too far to be useful.
Bad Routing to Enemy Wonders
Even though the AI it’s good at scouting, it never, to my knowledge, pursues the path of least resistance when targeting enemy Wonders. This is the main issue it has when dealing with them. AI armies can, and often do, encounter an enemy base and attack it, with all of its defences, when they could and should avoid it to reach the Wonder directly. Again, the good scouting the AI makes allows for a better planning of the route to the Wonder-ful target.
Battle Behaviour & Unit Management
This is the most noticeable subject. Regrettably, there are still a lot of deficiencies like passiveness, from occasionally but needlessly reforming formations while in battle or idling completely (especially siege units), or
inconsistent behaviour, from “cowardice” (retreating way too early) to heedlessly charging in under multiple Keeps without siege units (that may come on their own a minute later…). I witnessed ridiculous things like Trebuchets strolling by an enemy Keep and getting exploded by it. It got depressing after a while 
. Often enough, the AI looks like a panicking bronze player when managing troops.
Suicidal and Unfocused Raiding
When attacking the enemy base, the AI has an habit of running around under TC fire trying to kill things, especially Sheep (it hates Sheep), and even Ramz can fail to focus the TC and other defensive buildings. After that, Houses or Mills (to name a couple) can get more attention than production buildings or, more importantly, landmarks. The positive thing is that it’s actually idling someone, though there are better ways to do it.
Unafraid of Defences
It’s generally too careless of the enemy defences fire. Doesn’t stand off and wait for siege units to do their work, all-ins even when it has Trebuchets. May even have Bombards destroy a useless Mining Camp before the Keep that’s firing on the troops. There’s no such thing as priority.
They got Keep, I got Torch!
Attacks Keeps with infantry armies, even mostly comprised of ranged units, with no second thought. The funniest thing is that it eventually wins. That says a lot on the AI recklessness and the effectiveness of Keeps. They should build a few Rams and enter them.
No Walls, No Rams
It’s not exactly like that but Rams (and Cheirosiphons) are very rare. Considering how Rams have changed (and how humans love/loathe them) AI players should learn to use them more often, in particular when they have bonuses.
Doesn’t Drush
Some Civs have an innate talent for Dark Age Rushes, the AI completely ignores that. There should be a chance for the AI to carry out a drush, more so with those certain Civs.
Misses Raiding Opportunities
Often is so focused on hitting the enemy base, non necessarily successfully, that lets happy communities of enemy Villagers scattered around the map… happy. Should be capable of mustering raiding parties and hit those idyllic farming villages or trade routes. They get always scouted and are easy targets.
No Forward Bases
The AI doesn’t build forward bases to keep the pressure up, its armies always come from and return to the main base. Should known when and how to do it, could unlock a constant flow of reinforcement.
The French may happen to do it by accident if they place a Keep somewhere near an enemy base, as they will try to exploit their recruitment bonus.
No Aggressive Fortification
Kind of linked to the previous point, this could be a debatable one, but personally I expect the most intelligent AI to be capable of carrying out a strategy like this, if it sees it fit. Tower rushing, permanent siege with Outpost and/or walls enclosure or just a strong point to keep a foothold.
Relics in Battle
Sometimes the AI brings Relics in battle but it never ever even attempts at Conversion. Evidently the Monks carrying the Relics are inadvertently included in that army before they could deliver it. Relic carrying Monks should be prevented from being included in an army, unless there’s an idea of using Relics for Conversion, but the AI as to learn that beforehand.
Passive Siege Units
Not rarely, siege units will stand idle in the heat of battle. Other times, most often in an enemy base, they will pack-unpack-pack-unpack-pack-unpack… and so on, without attacking.
Siege Units don’t Actually Try to Outrange Defences
I saw them retreating at times, but usually never enough. Anyway, it matters little when the rest of your army is running in to get a boiling oil shower.
Siege Units Never Repaired in the Field
Instead of campaigning on until they break, sometimes siege units could get fixed by Villagers, before going up to the base (where they can actually get repaired).
Siege Units don’t Deploy by Default
When armies stop, their siege doesn’t deploy, with the result of not being able to fire until attacked (or not fire at all).
Bugs
Rarely gets stuck on Sacred Sites and doesn’t react to attacks on the main base.
Scouting
The AI is quite good at keeping an eye on the enemy base, but Scouts don’t follow enemy armies for interception or anything like that. In general I don’t believe there’s much consequence to what is being scouted, e.g. if a trade route, a lonely Monk or a TC is found, nothing happens.
Early Scouting
At the beginning, the Scout doesn’t make a full circle around the base to find the nearest Sheep and re-explores too much.
When looking for Sheep, if they get “spooked” by other Scouts, they can explore the same area again, instead of looking somewhere new.
Missed Sheep
They don’t always pick up Sheep in their LoS. If that’s what they’re looking for, they should!
Sheep Handover
AI doesn’t make Sheep handover between Scouts, it could help with the risk of losing them exploring the enemy base.
What you even came back for??
I witnessed Scouts get back to their base with 1-2 Sheep but they didn’t drop them off.
Foolish Scouts
Sometimes they get waaay to close to enemy armies. And die.
They can get too close to enemy defensive buildings even after Spyglass is researched. And die.
Unprofessional Scouts
The AI doesn’t Professional Scouts, rather runs Villagers out to the Deer (and then farms there). Surely the AI hasn’t multitasking problems with retrieving Deer with Scouts every now and then.
Late Game
Useless & Idle Villagers
Doesn’t have an efficient economic strategy for the late game, when resources are either depleted or inaccessible. Even if it does have one, the result is dozens of idle villagers kept there wasting population space. Should learn when its best to delete some of them, if they are useless. Also, should focus on heavy farming and especially trading, where possible.
Victory Strategy
Happens to get stuck in chasing only one Victory Condition, be it Landmarks, Sacred or Wonder, though the latter is harder to tell. This often results in very long stalemates, resolved only by sums of coincidences. The AI should be more flexible in pursuing various Victory Conditions, even at once, recognising when one isn’t taking anywhere at that moment.
Building Placement
Base Building
While it appears the AI does have a general idea of where the front of its base is, it isn’t consistent in the placement of its building. It has little concern where specific buildings are placed and many special ones are handled with the normal behaviour, thus wasting or limiting their bonuses. It’s usually compact and defensible, buildings tend to be grouped together but it lacks a bigger view, a real plan.
Landmarks
The most awkward placement choices are the Landmarks ones. They don’t seem to receive special consideration, can be placed on the front of the base, even if they’re economic (e.g. Grand Winery or Twin Minaret Medrese), or far far away (e.g. Palace of Swabia, King’s Palace or again Twin Minaret Medrese). Landmarks should be treated differently from normal buildings, accordingly to their possible special function, and placed in more defensible positions, when possible.
Useless Economic Buildings
Something that prevents the AI from achievement optimal base building is not knowing when it’s best to delete the useless economic buildings to free up space. This is particularly relevant in the case of Lumber Camps as several are made per woodline, taking up a lot of space, leading to debatable building placement, even affecting good Wonder placement. Inactive economic buildings could be considered as free space when trying to place a building, and deleted when the decision is made. A special check is needed for the Chinese, as they have to consider uncollected Taxes.
Safe TCs
Building TCs on resources around the map makes sense in the early to mid game, but when those are depleted the new ones should be mostly kept in the main base, with no more than a couple for backup. Relates to farming as well.
Housing
I noticed a tendency at getting housed in the transition to the Feudal Age.
Anyway, the AI is too strict on where to place Houses. Building in the base is not always possible but the AI forces it. This could also mean more time walking. Houses needed in a pinch should be placed the nearest to the tasked Villager.
Trade
Normally isn’t very good at it, never makes more than one Market so can’t properly trade boom. Trade routes are rarely optimal and many Civ specific bonuses are ignored (like Malians’, French’s, Mongols’). Even when lacking in Gold income there isn’t always a good response of training more Traders. Sometimes the AI seems to forget to rebuild the destroyed Market and all Trader go idle indefinitely.
Farming
Placement
Usually it’s pretty good but considering how easy it is to place farms for a human player, the AI ones should be mostly perfect. The bigger problems are another 2 though, linked together.
Isolated Farmlands
When collecting food around the map the AI places Mills, in the right places. Although in the late game, it uses those far and away Mills to farm, leaving farmers extremely vulnerable and even risking falling in a efficiency trap if the Mill is destroyed, as the farmer will keep farming there and going back and forth to the nearest (but hardly close) Mill/TC. A behaviour best to be avoided.
Lack of Dedicated Space for Farmlands
The AI avoids building new Mills for farming and uses the existing ones and TC as written above. A sounder tactic would be planning safe zones in the main base where to place extensive farmlands in the late game. Now it relies in building new TCs in basically random location and putting farms around it. Without TCs, there’s no farming. Actually it does make Mills near Farms, but it’s very rare.
The only exception are the English, they always keep one Mill in the base (and only one) to correctly exploit the farming bonus.
Farms built around unusable drop-off buildings
Happens in 2 cases: incomplete TCs or destroyed drop-off Landmark.
Sometimes the AI places a new TC, probably just for farming, and starts farming around it before it’s completed, causing long distance farming, not necessarily for a short while, since TC builders can be scattered by attackers and the completion suspended indefinitely.
In the second case, it’s usually worse. Repairing a destroyed Landmark takes a lot of time with only 2 Villagers (the AI never tasks more) and there’s no guarantee it gets through. The AI makes the mistake of not prioritizing the repair of the drop-off building, or doesn’t care for its existence in general.
Recovery
The AI doesn’t have an effective strategy for recovering from base destruction. This is in the scope of Team and FFA games, notably those without Landmarks Victory.
The AI doesn’t prioritize the safety of its remaining Villagers, instead focuses on creating an army if it has production available. Although not completely wrong it misses the idea of laying low and recovering. The AI should try to restart the Villager production as soon as possible, building a new and hidden TC and/or attempting to restore the capital one. Not having flexibility with farming makes recovering even harder, complicated by the fact that AI team mates won’t help with resources or repairs.
Generally behaves like a headless chicken, Villagers run around aimlessly or freeze on trivial stuff.
For instance, there was a Ottoman AI player left with 1 Villager. It had a Relic in a Mosque and a Market but all that Villager did was standing idle full of food, gathered from a Sheep, that she couldn’t drop off anymore. She should have repaired the Capital TC with wood acquired through Relic and Market. If she did that the AI player would’ve had a good chance of coming back, since it had several Military Schools still standing.
There’s a whole array of dysfunctional behaviours that hinder a successful recovery (ill-spending Wood, not re/building TCs first, not building drop-off buildings, going in dangerous areas…).
Doesn’t Migrate / No Backup Base
Although with the Landmarks system migrating can never be forever, in team games building a new base near an ally could turn the tides of the game. The AI should learn this, create a fully fledged base in a safe place from which restart production, before trying hard to rebuild the original one.
Forever Halted Repairs
Can happen that the repair of Landmarks isn’t resumed after a failed attempt (tasked Villagers died or Wood ended). This can be fatal.
Others
Excessively Delayed Textiles
The AI usually gets this tech very late in the game, after its done producing Villagers and despite losing many of them in attacks. I get delaying the tech but this is borderline nonsense.
