AI programming in general

[Max_Caine]

At the suggestion of Charon and Giis-Jan I have made this thread so people can discuss free AI programming in general, and in specific AI programming and Xenonauts. So have at it! 

[Charon]

So Chris was asking some AI questions in this thread https://www.goldhawkinteractive.com/forums/index.php?/topic/21567-official-ai-testing-suggestion-thread/ and its my favourite topic, but i thought there is a lot more to say and discuss about this than his request would allow. Seperating the wheat from the chaff so to speak.

AI programming is really my favourite topic, as i really think it makes or breaks these kind of games, and i see the most potent improvements that can be made for a game in this area.

So first we have to ask, ... what are we aiming for ? What is the kind of AI we want to have ? Generally speaking, we want the AI to be good and fun. Good in a skill way, that is we want to see and overcome something that understands and applies the rules of the game competently. Nobody feels satisfaction over beating the dumb kid, at best we feel quite bad, because A. we made somebody else feel bad and B. we didnt learn anything in the process. A Lose-Lose scenario so to speak.
On the other hand we also want the learning process to be fun. But what is "fun" ? What is fun for human beings ? Having a "fun" time usually means that your brain analyses and recognises a pattern which execution brings you closer towards your goal. You "understood" the rules and can manipulate them in your favour. Hence your brain rewards itself when it finds the execution of a pattern to be more efficient than other comparable processes. It says "Hey, this is the way to go, do more of that".
A few examples.

• An infant is trying to affect the world. Bigger impact means bigger reward. The mother builds up a block tower, the infant destroys it. The mother yells "oh no", and the infants brain fires off a big reward because it maxmially affected the world, and the people in it.
  The same applies when learning how to crawl. Do a lot of random behaviour, reward the successful ones, and you might one day have a very efficient set of skills. Everytime the infant tries to crawl, and makes some sort of progress, the brain fires off the reward structure which says "Maybe do more of that, you are getting closer to the goal of being able to move".
• When you indirectly throw a grenade and affect a target with the AOE behind an obstacle your brain goes "Hey, I just learned that i can indrectly affect a unit with an area of affect impact, without having a direct fire opportunity, and that brings me closer towards disabling that unit. Do more of that"
• When you are putting your sniper rifles in a more distant position your brain realises that "Hey, I can keep getting closer to the goal of keeping my units safe by putting them farther away, and the longer range of the sniper rifles still means i can affect the battlefield. Maybe do more of that."

Everytime your brain has this "AHA" moment, it rewards itself, and you feel good.

What your brain mainly does is pattern creation and recognition.

 

So, how do we get there ? The corresponding literature is very fascinating, just reading from Half Life to Halo AI development really makes you realise how much thought intelligent people put into their game trying to make their game as fun as possible.  Its definitely worth to read some of the developers roadmap diaries,m and just compare them to how you would do it yourself.
Unfortunately the statistical data is very clear. You cannot deduce, or induce what makes a game fun or not. Though we do have a proven concept on finding fun in a game: trial and error. Its not the fastest concept, but it works. In the same way that an infant does random behaviour, and rewards elements that bring it closer towards the goal, and it disregards elements that "just dont seem to do much". Half life had that problem that once they had their first prototype, and played it, they just found that "it wasnt really fun. So we had to scrap everything and start over.". They then succesfully applied trial and error to the development process, identifying and reinforcing elements they collectively felt were "fun", while dropping elements that nobody felt strongly over.

Lets now go one level deeper into AI development, and away from a more general approach.

When we talk about a good AI we have to ask how much information we want the AI being able to see, in comparasion to a human. My general oppinion is that the AI should be as humanlike as possible, that is it should have behaviour that a human with the same amount of information is more or less likely to come to as well. In one phrase, the behaviour of the AI should be reproducable for a human player. Questions that fall under that are

• should the AI know about sight ranges for units ? Is that information the player can play with as well ?
• what about armour values, should the AI know about them ? Does the player have any knowledge or interaction possibilities for that.
• what about items a unit is carrying, should the AI be allowed to know what a unit is carrying. Is that something a player can collect knowledge

The main point to set here is really to set the board as equally as possible, so that both parties - human players and bots - can make their decisions based on information that both parties could possess, in order for human players to look at the AI and say "Hey, thats smart, i should propably do that myself". That is propably the biggest performance hog though, as more complex behaviour leads to a slower game. A balance has to be struck here.

 

On the fun side we are looking at pattern recognition. One of the fun things the Half Life team discovered was getting killed by something that was out of scope of the players knowledge wasnt very fun. On the other hand if they get a demonstration of what not to do, and where the danger lies, the players felt more like that their failure was on them, instead of an unknown variable. So they saw a pattern, and the impact of that pattern, and could then either apply or avoid such behaviour. Combining that with the fact that each unit has its individual goal in order fullful the overaching goal of the mission we end up with archetypical behaviour for units. I think it would make a very fun game if the player would be able to observe and recognise archetypical behaviour in opponents units. Such roles could be:

• Spotters: Hide in places which have good sight cover, while also having good vision. Only engage when facing lone units, or when the overaching goal of the map cant be achieved by other means than acting yourself ( AKA everybody else is dead ).
• Aggressive: Care more about getting into places with good shooting abilities than having good cover opportunites.
• Close Range Berserk: Try to get into close range by all means possible, dont shoot until you have high hitting chances
• Ambusher Active: Try to stay undetected as long as possible while getting into the best firing position. Try to make very few shots, but with high hit propability.
• Defensive: Priorities position with the best cover, take many shots, even with low hit% chances. .The overall feeling should be a lot of danger, but nothing is really hitting.
• Ambusher Inactive: Try to set a trap where you can maximise the damage of reaction fire. This can mostly be utilised for units inside the UFO, but also for units setting a trap inside a house.
• Sniper: Prioritise being out of sight range, getting into full or half cover, and only engage when other units have sight on targets, unless everybody else is dead.
• Terror: Priorities actions that decrease morality for soldiers, like killing civilians.

The feeling of the player should be "Aha, i know that behaviour, thats a Spotter." or "That unit is definitely trying to ambush me".

 

Lets go one level deeper.

On the good side i definitely feel like the game could use something i call a sound ripple function. Basically anything that is "loud" or flashy should attract the attention of all kinds of units. The louder it is the more attention it should attract, up to alerting every unit on the map that something has happened there. This should definitely be moddable to allow people to create weapons or effects with are "louder" or more "silent" than others.
The way that i imagine a potential game is that i support Chris`s vision of having "patrol" behaviour for units. And depending on how "loud" or "silent" you operate on the map you can attract or avoid attention by other units. So depending on the map layout you could start with a scenario where nobody knows that you just arrived up to a scenario where every unit on the map gets a notification of your arrival. And everything inbetween. That could easily be manipulateable by a sound ripple tile placed ontop of the arrival place. If the concept is good, you could even expand on it by implementing a feature that lets players give the pilot directions in whether or not they want a stealth approach, or a flashy big-bangs landing. The concept is very flexible and expandable in its application.
The other thing i mechanically would suggest is that units can "hijack" other units for their purpose, and form temporary groups. For instance a sniper behaviour unit could hijack a defensive unit and override its behaviour to be a spotter, even though that would put the unit in more danger than the defensive behaviour would allow. That behaviour can last one or two rounds and then units go back to being "individuals". Another example would be an aggressive unit infecting other units to be aggressive too, and then they start an all out assault ( irrelevant of the success chances ).
Consequently we are approaching the suggestion that the AI needs to keep track of some gloval variable like [] how many friendly units are alive [] how many enemy units dp we know of [] how many enemy units do we suspect to be there [] etc ... . Kind of like a semi-global-commander for the AI.

 

Anything deeper propably goes into tangible algorithms and the perofrmance aspect of it.

So much for the theoretical approach. Lets get some practical examples.

 

---

 

First of all i really have to say that there are a lot of people better at programming an AI than me. @TacticalDragon has made a lot of improvements over the Xenonauts 1 AI, fixed a lot of bugs and layed the foundation for any bigger mod that X1 has to offer. I definitely think you should check out  XNT https://www.goldhawkinteractive.com/forums/index.php?/topic/11800-x108-xce-025hf-xnt-into-darkness-mod-v60-quothellgatequot/ here. A lot of his work got incorporated into X-Division and i would say that 45% of the AI code is his, while 45% is from drages, and i supplied the remaining 10%. Though i think he went professional, and does AI development on a more paid level now.

Everytime i look over the AI in X-Div i think "This is really how i want my players to experience the game". This is a more endgame-all-out-battle where the player gets dropped more or less in a very hot combat zone, with a lot of tactical options nearby. You can see a lot of archetypical behaviour here, units trying to assault and get as close as possbile, units being satisfied at rifle range, cannons and snipers behaviour. You can notice that the overall feeling of the fight is that its very dangerous for the player, but not a lot is dying. So you get a lot of adrenalin, but not as much of a "I messed up" feeling. Basically there isnt a lot of insta kill going on, and you can "feel" it when you get closer to more risky actions. The punishment is more gradual, than instantanious.

https://youtu.be/5w69Kmgin3E

[Max_Caine]

There is something I'd like to say about AI and available information. It is possible to make the AI seem like it's cheating, even when it isn't. Take an average map. Sprinkle enemies throughout the map so the population density for any particular map area is quite low. At the start of the enemy turn, tell each enemy to spin in a 360 circle before any enemy makes any other kind of move. Because the data collection points are scattered throughout the map, the AI is instantly aware of 60%-90% of the map (depending on map construction and enemy placement). With that information, it's then possible to logically deduce where the players troops are on a negative mask of the available information. Because the player doesn't know that the AI data collection points are spread throughout the map, it looks like the AI is cheating, even when it isn't. Data collection and data sharing for the AI is a contentious issue. Just how much do you "legally" allow? 

[Charon]

1 hour ago, Max_Caine said:

Just how much do you "legally" allow? 

Im not developing an AI so i am unaware of the practical issues taht can arise.

Though i am not sure i am seeing your problem. Your problem assumes that the AI is pre-emptively looking for the player without any tangible evidence. Given that this is a computer game, and map size affects perofrmance, and therefore space should get used sparsingly an intelligent AI can indeed backwards calculate where the player is most likely based on where he is not. There is an easy solution to this. Dont implement a function like that.

No really, just dont do it. I am on Chris´s side on having more or less patrol behaviour around the mapm unless there is tangible evidence of activity in the area. So for instance lets say you shoot a rocket into a building, that is loud, and the surrounding aliens might check out what noise that is, and therefore they deviate from their standard path a little bit, while still obeying their main role behaviour.

I would rather have aliens looking like they are cheating, but they are not, instead of aliens that dont look like they are cheating, but do. If the aliens behaviour is legit, its something that the human players will be able to copy at one point in their playthrough, and have fun, instead of having what looks like to be legit behaviour, but is actually impossible for the player to do.

Cheers <3

[Max_Caine]

The AI behaviour is legit. It's something that players can and do do during their turn. As the map is explored, the player can and does induce where aliens are based upon where the aliens have not been. The more of the map is explored, the tighter the ring forms around where aliens can be. The difference between player exploration and AI is that if aliens are dotted around the map to start with they get to explore the whole of the map more quickly than the player can. If the individual AI unit could not share information with other AI units this would be a moot point - they couldn't build a coherant picture of the world and react to it. However, if AI units couldn't share information, you'd get unusual and bizarre behaviour. AI could not, act as groups for example or respond effectively to player threats. 

If the AI is permitted to share information, then it would be probably be better to have aliens start at designated spawn points on the map. Then the AI's knowledge of the map would be more resitricted. 

[Charon]

I never suggested that everything that humans are capable of should also be something that the AI should be capable of.

I only suggested that everything the AI is capable of is something humans should be able to do as well.

 

This is a game. Humans play games in order to train. But in order to play you have to limit yourself to a ruleset. "Anything goes" is not a ruleset.

If you train martial arts you might one day train with a wooden dummy. Now clearly a wooden dummy does not possess the same capabilities as a human. Yet you can use it to train your body and technique. For that it is not necessary that you are both on equal terms.

A game is a wooden dummy. You can use it to train yourself, push yourself beyond limits, experience new hights and/ or entertain yourself.

 

Though i think we are moving a bit deep into game design with this.

[Bobit]

It's a good question because other games just don't have shared vision, or have vision only shared with the squad (pods in FiraxisCOM, spotter units giving sniper units LoS in OpenXCOM, to a lesser degree the cheaty intelligence stat which lets AI track any unit they spotted for a few turns). Because "good enough" really is okay when it comes to unspotted AI behavior, but there's definitely a lot of room for improvement.

AI can see 60-90% of the map in this game, unlike the player, because there are so many of them, they all have extended LoS, and their starting positions are spread out. But if they did not have extended LoS, as in XCOMFiles where most foes are human, this would not be the case. Also they can only do this if they spread out, then they are easier to deal with perhaps.

About the "ripple", something simple like a "shout" ability that alerts all nearby aliens (as aliens do not share vision by default) does great in Long War, Cogmind, especially if the player can visualize it. Perhaps the sound should be sourced from a random location within 3 squares, so that the AI does not know exactly where it comes from.

I definitely like the idea of dynamically-formed response squads, but it would need to be simple, maybe just this: join the squad of whoever is close, sharing intel and moving together (i.e. both prefer to chase 1 recently-spotted but currently unspotted target rather than split up), and share the "sneakyAi" variable from OpenXCOM which basically means they heavily prioritize blocking enemy LoS at end of turn even if it means they don't get to fire (if one is not sneaking, none will sneak as all but one sneaking would just result in that one tanking all the fire) (of course Xenonauts has less LoS-blocking cover and more accuracy-reducing cover, so it's not quite so simple). I wouldn't expect the AI to know how long to "wait" before an "all-out assault", or to refuse to fire for fear it would reveal their position, even players often take the greedy approach so you can't expect the AI to know when not to be greedy.

Edited April 21, 2021 by Bobit

[Chris]

Some long posts here so just a quick response from me here. I think the most important thing for the AI is just to punish you when you make a clear mistake. There's nothing that's quite as immersion breaking to end up in a situation where you have a soldier stood out in the open in front of an alien that is clearly about to get killed ... and then the AI runs away or does something else instead of murdering your solider (or whatever incredibly obvious action you were expecting it to do).

That's happened to me a few times in the various XCOMs and although in some ways it is a relief (as I play Iron Man and don't want to lose soldiers) but it does make me like the game less overall. What's the point of playing a tense tactical game if the AI doesn't punish you for making a mistake? Things suddenly become much less tense.

In most cases the player doesn't actually give the AI much of a chance to show what it can do in this sort of game, because most aliens get killed the turn they are sighted so don't really have a chance to act (except maybe overwatch firing), so the AI really has to make the most of any opportunities the player gives it.

There's lots more complicated stuff you can do with AI, of course. But that stuff is all icing on the cake rather than the cake itself in my opinion!

[Charon]

10 hours ago, Chris said:

Some long posts here so just a quick response from me here. I think the most important thing for the AI is just to punish you when you make a clear mistake. There's nothing that's quite as immersion breaking to end up in a situation where you have a soldier stood out in the open in front of an alien that is clearly about to get killed ... and then the AI runs away or does something else instead of murdering your solider (or whatever incredibly obvious action you were expecting it to do).

That's happened to me a few times in the various XCOMs and although in some ways it is a relief (as I play Iron Man and don't want to lose soldiers) but it does make me like the game less overall. What's the point of playing a tense tactical game if the AI doesn't punish you for making a mistake? Things suddenly become much less tense.

In most cases the player doesn't actually give the AI much of a chance to show what it can do in this sort of game, because most aliens get killed the turn they are sighted so don't really have a chance to act (except maybe overwatch firing), so the AI really has to make the most of any opportunities the player gives it.

There's lots more complicated stuff you can do with AI, of course. But that stuff is all icing on the cake rather than the cake itself in my opinion!

Some solid oppinion there :thumbsup: !

I get what you mean. The value of a players action is depending on keeping their soldiers alive. If the AI doesnt consistently punish your mistakes it devalues the players actions, because it puts part of the danger on a dice roll ( the dumb behaviour ), instead of making the player feel like they are responsible for each and every action they take.

[Charon]

I do support patrol behaviour, but i really think that putting patrol behaviour points on the map is problematic, and will become a double edged sword in the future. The appeal for games like Xenonauts for me is that you can load up a map, throw in some units, and you get a different fight everytime. And if you get a different map, or a procedurally generated one even better. The advantages of map patrol points is that developers can do some nice stuff with it. The drawback is it hurts replayability of a map if you know exactly where you can expect units. That might not be obvious on your first playthrough, but can get old very fast.

So i would really like to have general-purpose AI that you can drop into the map and just expect them to do reasonable well, with patrol behaviour applied ontop of it, instead of mainly patrol behaviour, with a minor layer of general abilities.

[Dagar]

I agree to the general points here: AI in games should be fun (as in an engaging, interesting opponent) and good (as in challenging, making good decisions within the decision space given). To this I would maybe add that in games, fun is always more important than good. If there is a strategy that makes the AI very good but that is "uncounterable" by the player, that would not be very fun.

One thing that in my opinion is missing from the discussion here is that ideally, it should react to the player, as in adapting to counter the player's play style. This is more on the focus of the overall campaign rather than a single mission and could take multiple facets. For one, the AI tactics could change, e.g. in the form of the balance of different behaviour archetypes being adapted. If the player likes to blow up cover and snipe everything from a distance, the prowling, sneaking archetype might become more pronounced. The other way to adapt is what Phoenix Point tried with their evolution of enemies according to what happens in missions. That does not have to be genetic adaptation, but can also be in the form of new equipment. If the player likes to rush in with shields and shotguns, it would feel great if the aliens could try to cover their approach ways with incendiary grenades and the use of a better suppression weapon.

The most important thing I feel there is to consider here as a developer is that the change in behaviour should be gradual, so that the player has the chance to observe the new strategy and counter it herself. And if the system regresses to the same optimum regardless of the player, that would most likely be the fault of constraints or the overall game system being flawed.

 

As for the waypoint system, it is pretty easy, from a programming perspective, to make it unpredictable even with the same map blocks, you can distribute many possibly traversable waypoints over each element, choose for a particular patrolling enemy or group a rough patrol path over the map (most likely roughly circular or elliptic) and randomly select waypoints in the vicinity to that rough path, with the closer waypoints being more likely to be chosen.

Finally, for what the AI should know and whether it should be more than the player can know, I'd say keep to what is most fun but communicate the AI's capabilities to the player. An enemy that can look through walls can still be fun, if the player knows that and can act accordingly.

[Bobit]

14 hours ago, Dagar said:

 

One thing that in my opinion is missing from the discussion here is that ideally, it should react to the player, as in adapting to counter the player's play style. This is more on the focus of the overall campaign rather than a single mission and could take multiple facets. For one, the AI tactics could change, e.g. in the form of the balance of different behaviour archetypes being adapted. If the player likes to blow up cover and snipe everything from a distance, the prowling, sneaking archetype might become more pronounced. The other way to adapt is what Phoenix Point tried with their evolution of enemies according to what happens in missions. That does not have to be genetic adaptation, but can also be in the form of new equipment. If the player likes to rush in with shields and shotguns, it would feel great if the aliens could try to cover their approach ways with incendiary grenades and the use of a better suppression weapon. 

The most important thing I feel there is to consider here as a developer is that the change in behaviour should be gradual, so that the player has the chance to observe the new strategy and counter it herself. And if the system regresses to the same optimum regardless of the player, that would most likely be the fault of constraints or the overall game system being flawed. 

 

I actually feel WoTC did a much better job of "adaptive AI" than Phoenix Point. Rather than changing AI based on player kill statistics, simply make the bosses (or areas, units, factions, whatever) that win even more important to beat. If the player tries extra hard to make the stronger enemies lose, well at least that is thematic and often makes for high-risk high-reward tactics as those good bosses are the most dangerous, so even when the system fails to predict what will make the game harder for the player, it still succeeds in making the game more interesting for the player.

It's a very common philosophy too. Arguably any 4X has it (Long War 2 geoscape being such a 4X) as certain factions/areas get overrun it is more critical to defend against them. Games with difficult tech trees like X-Division have it as well, since it becomes more critical to capture a certain kind of unit the longer you wait, and capturing those units generally gives tools to counter them. XCOMFiles does too, as capturing a leader of a certain faction will allow you to perform research that unlocks higher-level missions of that faction and removes lower-level missions.

Phoenix Point and MGSV are the only games that use kill statistics and in both cases it feels really forced, slow or not.

Isn't the whole appeal of XCOM the tech race after all? So why not use a solution that focuses on the tech race?

Edited April 23, 2021 by Bobit

[Charon]

@BobitI feel like you are making some good points.

I like games that evolve around finding a strategy A, that the game then counters with a Counter A, for which the player than finds a Strategy B, that the game then again counters with a Counter B, all the while Strategy A still exists as a possibility. And so on.

X-Division is propably developed with the idea in mind that that for any given problem there exists a tool which has the appropriate answer to that problem. The tool propably isnt directly available to you at the moment, but if you look for it you will propably find it. Some tools for problems get unlocked while you overcome the problem with sub optimal equipment. For instance beating interceptor class UFOs is usually a big problem early on, but as soon as you successfully overcame that obstacle the first time you will be able research a tool that makes achieving success over the problem a lot easier in the future. Sometimes a tool is hidden in a different place than the problem is. If you look into Caesan technology you will get energy weaponry, which is very good against sebillians. Sebillian technology provides kinetic weaponry, which in turn is very good against caesan and andron drones. And so on and so forth. The point of this is to make the discovery process as interesting as possible, and to reward the player to engage in the content of the game.

I always have to think about Zelda games when this point comes up. Zelda games are basically a game about collecting keycards to get access to be able to collect the next keycard. Daisy chain enough of these keycards and you get a game. It leads to a very enjoyable discovery process, and the hunt for the next tool to solve the next problem.

Another propable goal of X-Division is periodisation. Thats a concept taken from Mark Laurels Bodyweight Training guide. He found the best results for training humans in periodising the intensity and repetition of the exercises. That is you change your training regiment between high intensity and low repetition with low intensity and high repetition. The human body can not operate on peak condition all the time, but periodisation of peak conditions gets the most out of the human potential. And so the game tries to throw few missions with high intensity, and many missions with low intensity at the player, enabling the player to choose their own periodisation (mostly ). One of the flaws of later game content is that it overburdens the player with the amount of high intensity missions in order to make meaningfull progress in the game. The game doesnt let the player relax with low intensity missions, while still partly making meaningful progress.

 

So i think you can say that the most enjoyable "adaptive AI" has something to do with what kind of tools the player has access to at any given moment, and what kind of problems the AI can throw at the player at any given moment, and less to do with different AI settings in Ground Combat. You basically change the tools the AI has access to, but not the programming behind it.

[Dagar]

On 4/23/2021 at 11:31 PM, Bobit said:

I actually feel WoTC did a much better job of "adaptive AI" than Phoenix Point. Rather than changing AI based on player kill statistics, simply make the bosses (or areas, units, factions, whatever) that win even more important to beat. If the player tries extra hard to make the stronger enemies lose, well at least that is thematic and often makes for high-risk high-reward tactics as those good bosses are the most dangerous, so even when the system fails to predict what will make the game harder for the player, it still succeeds in making the game more interesting for the player.

It's a very common philosophy too. Arguably any 4X has it (Long War 2 geoscape being such a 4X) as certain factions/areas get overrun it is more critical to defend against them. Games with difficult tech trees like X-Division have it as well, since it becomes more critical to capture a certain kind of unit the longer you wait, and capturing those units generally gives tools to counter them. XCOMFiles does too, as capturing a leader of a certain faction will allow you to perform research that unlocks higher-level missions of that faction and removes lower-level missions.

Phoenix Point and MGSV are the only games that use kill statistics and in both cases it feels really forced, slow or not.

Isn't the whole appeal of XCOM the tech race after all? So why not use a solution that focuses on the tech race?

I have played neither PP nor WotC (not buying FiraXCOM ever again after the disaster that was XCOM 2 for me) or Long War 2 (but 1 I have played), but I feel that what you pointed out here is not exactly what I'd like to happen in a game. X-Division, since Charon brought it up, also does not exactly do that for me.

Yes, you as player have access to a tech tree and it enables you to alter your tactics and perform operations you previously could not. But the enemy in all these games is pretty static. Yes, they gain stuff over time, sometimes even according to their success over you, but never in response to your actions. Basically, the aliens escalate pretty much with or without your interference. What they encounter on the battlefield does not matter, at most the outcome of a battle or of the overall campaign determines their progress.

What I'd like to see is both their equipment and their tactics or maybe even strategy change according to what you do. Ideally not only on the battlefield, but also in the air, with your bases and with your tech.

For example, in the beginning they only send scout UFOs to see what humanity is up to, and to scout areas for bases, bombing and terror attacks and so on. If you shoot down 90% of these, they should start scouting with escorted scouts or with bigger UFOs (that is the equipment side) or overwhelm an area with multiple scouts that you cannot all shoot down, or do multiple very short "dip into the atmosphere for a few minutes and get out quick" missions. Instead, what we get is heavier scout UFOs at some point, irrespective to our success in preventing scout missions. Instead, we get steady progression into generally stronger, bigger, better enemies (Long War, FiraXCOM) or leaps in progression (X-Division), which are disconnected from what happened before.

Maybe some game does this, but if so, they fail to properly communicate that this is happening. I brought up PP, because from this game at least I know that the ground units "evolve" according to what you do (even though it seems to be a pretty abusable, broken system).

I want an enemy that rushes AoE weapons, and concentrates on area denial and flanking hit-and-run type enemies if I always waltz over the map with a squad of close range shield bearers. I want the enemy to emphasize cover and teleporting into my backline if I have a strong sniper-and-scout combo going for some missions now.

[Bobit]

@Dagar

The big hook of WotC is that there are 3 re-appearing bosses with randomized traits, and they get more of such traits as they defeat you and you fail to defeat them. So the ones with strong traits or traits that counter you inevitably become stronger and stronger. Meanwhile on the XCOM side you get a few levellable hero units. I am not a huge fan of vanilla but other than the tedious zombie missions and the usual balance problems everything in this expansion was perfect.

I do think the kill-statistics based AI is extremely suitable for the air game, as a non-interceptor UFO's success is almost entirely based on whether it is killed, and it's not easy to simply change your plane's equipment, and compositions/roles are not so complex. For ground, well maybe it's better to reduce the complexity first, for example by considering a boss's or faction's success.

Edited April 25, 2021 by Bobit