Even on Humankind difficulty the AI is not even close to being able to keep pace with a human player. These are the statistics charts from a game with Egyptians, Maya, Khmer, Mughals, and Italians:

Here is the corresponding save file: Italians Turn 169.ctr

So the game was already decided before turn 100. In terms of fame, the AI actually never surpassed me.

My general strategy with this game was to pick builder cultures and focus heavily on production so that I can finish a district ever 1-2 turns in each city. Then I was ignoring stability completely until I got the Italians which thanks to the high production were able to build commons quarters then so quickly that a stability target in the negatives were fixable before the cities hit the 30% threshold. Yes that is a somewhat advanced min-maxing strategy. But even with clever min-maxing I still expect the AI on the hardest difficulty to not appear like an almost flat graph compared to exponential growth of the human player.

When I look at the districts that the AI builds it looks like they build many commons quarters in relation to small amount of overall districts that they have. I assume that is probably the reason why the AI is not doing so well even though it gets a lot of bonuses. Maybe the stability bonuses for the AI on higher difficulties need to be buffed a lot more so that they can build more districts without worrying too much about commons quarters. I don't really know in what order the AI builds its districts, units, and buildings, but I assume that maybe the AI is not really prioritizing the right build order. I personally try to focus on production first, then food, then science, then gold. I also try to have as few units as possible. I try to keep a small amount of units for defense in border cities if there is an aggressive neighbor and sometimes use units as stability boost if a city temporarily needs a bit more stability. I am not sure if that is also what the AI does. It is hard to tell due to fog of war.

Another thing that puzzles me sometimes is that when the AI declared war on me, it sometimes sends some weak scout units in my territory, which then retreat when I attack, but since they are in my territory I would kill them sooner or later anyways. So that is just free war score that he AI gives away. I remember someone saying that humankind AI does not suffer from fog of war. If that is so, then the units should see that I have other units nearby and that a retreat will just prolong the defeat and help me win the overall war. In those cases the AI should strategically sacrifice their units in order to maximize the war score. But first of all, the AI should never send weak scout into enemy territory when they are at war.

Great job on the AI!  It was a huge improvement over the previous OpenDev, or at least seemed so to me.  I felt like I could pretty reliably predict what the AI would do, simply because it would take the most advantageous move.  Fights were significantly less one-sided than previously, and even on a 5v3 basis I'd still have to be cautious or lose a unit.  I only had one case that really felt off, and that was when the AI moved their archers in front of their warriors to shoot my units, which let me melee attack them from uphill.  I'm thinking that there may have either been some sight-related issues, they prioritized shooting my archers, OR they determined that losing archer health was more important than losing their warrior's health (in an uphill, losing fight).  Obviously it's hard to say (especially without pictures, sorry!) but there were several cases where I chose to "tank" with my archers on melee attacks to prevent losing/weakening my melee units, so I could see it as a viable strategy overall.  That case was probably their only significant blunder.

I do have a couple extra notes on potential AI improvements, although honestly I'd prefer them as higher-difficulty choices (for the sake of difficulty scaling).  The main one is that the AI does not seem to prioritize hitting weakened units, nor on ganging up on a single unit.  It was usually pretty simple to retreat near-death units to the back to avoid death, and often even enemy archers would avoid killing them.  There is some merit in the AI's decision to avoid those fights and focus on weakening the overall force, but it would often mean my units live to fight another day.  Similarly, while the AI loves being offered an archer to melee attack, they would often not focus multiple units on them.  This was especially detrimental to them, as I could use archers to split the damage where otherwise one archer could have completely died.

Last AI note:  their overall army management may need some work.  The AI seems to favor a divide-and-conquer approach most of the time, which is honestly terrifying and is often effective if they have the units to spare.  However, this also means that every time they fight, they are likely to lose the battle (at the cost of potentially ransacking unchallenged).  Mid-war, I could often attack a new army each turn with overwhelming force, and that means their war support would crumble.  Since war support is not based on the number or quality of troops lost, it was fairly easy to win a war by picking off small armies so long as I could avoid losing a city (through ransacking/rushed enemy units).  I don't think their default approach should be removed, just maybe re-evaluated.  My first major war I only won because I could funnel invaders to a two-lane pass and alternate hitting them on either side, which meant that they'd have been more successful with one army of 5 than two armies of 2 and 3.

Additional separate note, the current difficulty scaling is harsh.  I started trying to play on "Humankind" difficulty (as last opendev was way too easy), and got completely and regularly wrecked.  I don't know how the AI is generating so many units (warriors, spearmen, and archers!) when I don't have anywhere close to the population (much less production and turns) in the first 30 or so turns.  My guess is that higher difficulties give AI free units, population bonuses, and/or production bonuses (like many other strategy games), but this can give the impression of "cheating" if the AI is smart enough.  Typically, giving free units to AI as a difficulty mechanic can be really discouraging and lead to a broken early game (looking at you "Civ", with your settlers!), so if that is being done then just be cautious.  I would MUCH prefer a scaling AI intelligence over scaling global benefits, but I recognize that in a complex game such as this that is hard.  I ended up wasting the majority of this opendev trying and failing to play on "Humankind" difficulty, which was rather demoralizing.  I'm not saying it shouldn't be hard (obviously it should!) just be careful that you don't expedite the "exponential growth curve" of the game too much by giving the AI too many free benefits at harder difficulties.  I've learned my lesson, and I'll save the hardest difficulty for when I've really understood all mechanics of the game.  Looking forward to August!

Here are a few points where the AI could be improved:

Outpost/city placement: 

The Ai doesn't seem to be able to calculate a fallback option when the tile it wanted to use is occupied. This leads to an exploit where you can just block the AI's expansion by placing a single unit on the recommended city spot of a territory.

a few potential solutions:

1. Have the AI attack the unit if it seemed to be camped in place for more than a few turns (of course it would evaluate the chance of winning, etc.).
2. Have the AI prioritize claiming territories above placing the "best" outpost. Especially since it should be able to move the outpost later.

1. Does the player have armies near that outpost and if so, how strong.
2. How far is the new territory from the player's city (the further the city the less chance it has of being attacked by the player)
3. Does the territory contain resources? Does the AI already have access to those resources?
4. Is the player already at war with other AI?

I played on difficulty (civilization or humankind not quiet sure) difficulty and i am really a noob at the game(Civilization i needed like 100 hours I really understood every gameconcept to its full potential). And i fear that it is maybe a little too easy in the end. Lets say it i survived, not more not less but they only built old units despite me having very bad technolgy. I was always behind the era and didnt even reached idustrial in my game because i wasnt focused on tech but survivng wars(even won the one, it was epic) but maybe a little to easy for the highest difficulty.

(I have two games under my belt: the first on nation difficulty, the second on empire, both with the closed beta.)

"Does the AI play in a way that feels consistent with their Archetypes and Biases seen in the diplomacy screen?"

Yes and no. On the surface yes, the AI often does something which resembles their indicated playstyle, however it is often just an illusion coming from the voice acting. In the long run - especially without sound and towards the end - they play similarly. Only the hothead traitor does something extremely stupid time to time to prove its worth to the player regardless of the broad situation in the game.

"How did the AI evolve throughout the game? Did it feel equally challenging at the start and end of your game? Does increasing the difficulty increase the challenge as much as you expect?"

In the first 50 turns they are quite challenging regardless of the difficulty (declaring sensible wars, making good moves, participating in diplomacy etc.). However after this initial flaring they are fading away from the game very quickly. The usual issue of 4Xs. For example look at their fame progress. In my first game i won with around 11000 fame. Aminata came second with around 7000, and Isabella (my vassal btw) was at third with 4500. Okwaho was dead, Temulin was Aminata's vassal, and Yanga, Vlad and Keita were insignificant. The main issue here that basically this was the state of the game from turn 75. (I had - i don't know - 3000 fame, Amanita had 4200, Isabella had 3500 etc.) My second play through was interrupted at the turn 180-ish. I had 14.5K fame, Amanita had around 7K, Isabella and Okwaho with around 3-4K. Vlad was dead, Keita was my vassal, Yanga and Temulin was insignificant. And again this was the situation around turn 70 (only i had much less fame of course). On paper they were in the early modern to contemporary ages towards the end of the game, in fact however their present on the map was medieval at best. There are a lot of issues regarding the AI, however you should not lost in the details but find the main issue which is in the way of a good AI - and this is a problem of all the games in the genre. One more comment describing this whole situation: after around turn 100 i was literally avoiding and thus regretting to conquer any territory from the AI due to their cities smeared and snotty state.

My assesment: the AI's evaluation logic - if there is any at all - is 1) the same in all eras and with all cultures with only minimal (mostly cosmetic) fine tuning, 2) based on limited facts about the gmae in progress without any kind of TEMPORAL understanding. This creates a barrier which hindrances their ability to act upon the real situation of the current gamestate. One example which shows this glaring problem: You have a lot of luxurious resources they have a few. They want to trade with you, you counter their offer, they refuse to pay. (Always, which is another issue...) Next turn. The VERY SAME thing happens (no matter who they are). Next turn. Same thing. Next turn. Same thing. Rinse and repeat till forever or until some kind of violence happens between us which literally forbids them to make the same proposal. More over i have a hunch that the developers and/or the implementators do not have the necessary mathematically translated knowledge about how the game works, thus they have an impossible job to do. That's why you need to give the AI odds to increase the difficulty, which leads me to my final (view)point here.

Note on the difficulty settings: giving odds is not equal to raising the difficulty. I had beaten 4X games on "endless", "deity" etc. difficulty settings but i promised to myself i will never again touch those. The main reason is twofold: 1) the gameplay is exactly the same difficulty-wise as in any other setting, 2) these higher difficulties basically takes away gameplay mechanics and possibilities from the players which are in fact just sucking out the enjoyment from the game. On top of that i do not have any kind of achievement feeling from them because the only feeling they cause that i should petty those fools who were unable to win with those odds. I do not mind if you make one "masochistic-egotrip" setting with odds to those who fancy that kind of stuff, however i very much like to see at last real difficulty settings in a 4X game. To understand my view here let me illustrate this whole issue with a chess parallel. I play against - let's say - Hikaru Nakamura who is giving me 2 rooks, 2 bishops, 4 pawns and a knight odds to "equalize" the situation. Can i beat Hikaru with those? Maybe. Is that mean that is "challenging and difficult" for Hikaru to play this? Not at all. What should he do if he wants a difficult challenge? Play Magnus, Levon, Wesley etc. instead of me...

"Does the AI commit any grave strategic missteps? What is the AI doing and what do you think it should do? (Please provide saves if you can.)"

Yes. A lot. See the previous point.

"How does the AI perform in battle? Does it make any obvious mistakes it should avoid? What would you have done in that situation? (Please provide a save if you can)"

It is not bad, definitely better than in endless legend. However:

Tactical blunders:

1. Suicidal tendency instead of defending - in other words: it cannot grasp on the idea of winning a battle with defending their flag. (However it is partly because its inability to keep up techwise...)

2. The utter inability to understand and prevent the opponent's tactics. In other words there is no (or only very basic) threat assessment in the logic of target selection and move orders. (Try this: commence a siege with instant assault with only gunner/range cavalry units against a city which has mountain range (or other obstructions) in the way on the battlefield. In most cases with this kind of setup you should not win the battle regardless of the technological difference - you need at least one siege weapon. However the AI very kindly sacrifices their units one by one by moving them towards my units' range inside the castle instead of hiding behind closed doors far away from the pointy ends.)

3. The AI usually do not kill off units. It prefers targeting specific spot instead of units. (Try this: have a maimed unit at the beginning of your turn which was targeted in the previous one. Move the maimed one to a better spot (high ground for example) but inside the range of the opponent's units, and move a fresh unit to that spot. The AI will show off its skills...)

4. The AI does not recognize and therefore not use the edge of the battlefield as natural barrier (for example to defend its ranged unit).

However i have to admit that i have witnessed the most devious and resourceful tactic i have ever seen in 4X games form an AI: i had horsemen, the AI had a few low-tech closed combat unit and a few low-tech anti-cavalry. The surrounding was mountainous with lots of elevation. In line of sight there was a chargable closed combat unit with only one path leading towards it. I went for the kill. The ac units were hidden right next to the path in an elevated position waiting for the bite. I was in awe...

Tactical-logistical blunders:

1. The AI often wanders towards the player's territory with one army instead of keeping multiple armies together close to the frontline to be able to benefit from reinforcements. It happens more often if the AI has a superior unit. That being said i had seen also the usage of hiding reinforcements at the beginning of the game, but not in later periods.

2. Army composition: the AI either has a single recon unit (a horseman for example), or a group filled only with its best unit. (immortal, hunnic or mongol cavalry etc.) It should at least use combined arms for main armies (slow but effective), raiding parties with mounted units, ships and mounted unit as a recon and/or expansional force, and hired mercenaries as cannonfodders.

There are a lot more to discuss, but that is enough for now.

ritchiaro wrote:

How does the AI perform in battle? Does it make any obvious mistakes it should avoid? What would you have done in that situation?

-> The AI often strongly underestimates the advantage of playing defensively in battles. Especially in cases where I put a city under siege and attack the city, the AI often leaves the defensive place to come out an attack me. This has happened to me in many different cases and I also saw it happen to many streamers (even when the attack force only consists of horse units which are unable to pose a real threat for the city). So in those cases the AI often makes the huge mistake to attack. In many cases that means that I can win offensive battles without having to attack with a single unit, even on Humankind difficulty. That feels really wrong, because the AI should only attack if such an attack gives them an advantage. If an attack results in a disadvantage, the AI should not attack.

FatRabbit wrote:

1. Suicidal tendency instead of defending - in other words: it cannot grasp on the idea of winning a battle with defending their flag. (However it is partly because its inability to keep up techwise...)

That's strange. In my game, green sortied, then didn't actually attack my units, despite me only having 2 heavily damaged scouts, instead just standing around until the turns ran out, then immediately sortieing again.

Undernier wrote:

Yielding territories and outposts through diplomacy:

I have noticed that, early game, it is fairly easy to have an AI just give you a territory that it just claimed near your own territory.

Here are a few things that the AI could take into account to make a better decision:

1. Does the player have armies near that outpost and if so, how strong.
2. How far is the new territory from the player's city (the further the city the less chance it has of being attacked by the player)
3. Does the territory contain resources? Does the AI already have access to those resources?
4. Is the player already at war with other AI?

Based on those the AI could then try to trade the outpost for money or even inspiration (as it did have to invest some of it to create the outpost) rather than just accept or refuse the grievance.

I have noticed the same thing! Great suggestions. I'd like to add that before founding an outpost, the AI should potentially also consider if it would yield it to the player if the player asked for it. And if so, prioritize founding outposts which they would not yield to the player instead to use their resources effectively.

I think the general theme of your post is that the AI is really bad at adjusting their strategy if something is not going according to their plan, which is also my observation. Another example for this is that it is easy to box in the AI and therefore preventing it from expanding. In the closed beta settling strategically to the south and expanding south quite aggressively boxed in green significantly. The AI then just tries to look for other land, but either yields it to the player or looses it to war since the new cities are not defend-able. The thing that surprised me was that even if the green AI picked the Huns, the AI did only go to war with me in a few cases when it had a higher war support of felt that it was stronger than me, which even on humankind difficulty actually did not happen often probably because the AI was quickly boxed in.

The suggestion here is: When the AI's strategy does not work (boxed in, no resources for their plan, ...) it needs to think of alternative plans and strongly prioritize then (e.g. only researching military techs, building up a very strong army, finishing for war support and making a lot of demands, ...).

Here are some ideas of how to implement this AI:

- The AI should have a set of goals (e.g. expansion, access to resources, next culture, getting era stars, ...)

- Each goal should have an importance weight that corresponds to how much value the AI plans to get out of this goal

- If the AI is no longer able to reach the goal (e.g. boxed in, another player chose that culture, resource is not available in their territory), the AI should commit equivalent-value resources into an alternative plan (e.g. going to war, trading, staying in the era for longer or advancing quicker, protecting outposts, ...)

- The AI should regularly check what could prevent them from reaching the high-value goals and commit a reasonable amount of resources to protect the high-value goal (e.g. of expansion is very important due to the map and a close neighbor, the AI must protect their outposts from being ransacked with stationing a few units on the outpost, since ransacking outposts is currently a big weak point of the AI)

- If a goal is very high value and unlikely to be fulfilled (e.g. not a lot of land, but needs a strategic, needs a certain culture for synergy but is behind in era progression) the AI needs to already commit a certain amount of resources to an alternative plan.

So overall, the AI should get a better understanding of the concepts of value of a goal, its risk, and alternatives.

Some other more concrete suggestions that fit into the risk / goal / alternative framework above and which are important to make the AI competitive on higher difficulties:

- If the neighbor is building up an army and making demands (risk to their goal of keeping their cities), the AI needs to invest strongly into building a strong defense

- If the neighbor is ahead technologically and being aggressive (risk to the goal of keeping their cities), the AI needs to prioritize science stronger

- If the AI actually ends up being boxed in (goal of expansion) and can't go to war because it can't catch up with the player (risk of the alternative), the AI needs to strongly prioritize naval game-play to discover new land (another alternative). Is is basically what happened in Europe in the real world, so the AI needs to have a better understanding of the need to look for other land and use their resources to getting there.

- If the AI needs resources for their EU and plan to fight other player because they are militarist, then the AI needs to make an alliance with a player who has access to the resource. Strategic alliances is something the AI currently does not really work towards it seems.

I hope this was helpful and helped the devs to understand what a human would do in these situations and what are the risks we as humans plan for so that the AI can also do similar things. The great thing about an AI is that it is easier for an AI to periodically check their goal, risks to the goals, and alternatives due to more computation power. So if the AI would have a stronger emphasize on strategic planning, risk evaluation, and the concept of alternative plans, the AI should be able to keep up with the player even better.

• Does the AI play in a way that feels consistent with their Archetypes and Biases seen in the diplomacy screen?

Sometimes. Some archetypes are harder to notice then other.

• How did the AI evolve throughout the game? Did it feel equally challenging at the start and end of your game? Does increasing the difficulty increase the challenge as much as you expect?

It was way too hard on Nation difficulty and way to docile on metropolis, because you limited expansion too far. 

• Does the AI commit any grave strategic missteps? What is the AI doing and what do you think it should do? (Please provide saves if you can.)

Aside from weird balancing problems with difficulty levels - none that I've noticed.

• How does the AI perform in battle? Does it make any obvious mistakes it should avoid? What would you have done in that situation? (Please provide a save if you can)

Well, when I attacked one city in late game, it sent all peasants onto my Heavy Machine Gun. Needless to say, they were one-shot before they could even step near and deal any kind of damage.

Overall the AI felt much better than previous opendev, in general it just felt smarter but also they seemed to stick more to their personality traits

Idk if this falls into AI problem or not. On replay I see they advance through the next era with only 6 stars. Is this difficulty handicap? I played on Civilization difficulty. If yes, this doesn't make the AI more challenging, it does the contrary: It makes them lose fame score very quick. The only one keeping up just the Black one with a lot of stars on each era.

I know there's an avatar's trait/personality on advancing ASAP, but you shouldn't make it advance on 6 stars. And it shouldn't be too common. From 8 players, only me and the Black were collecting fame on each era, the other 6 definitely didn't.