Better AI in games

[Sande]

Very few people defined what they mean by "good" AI. Not to mention it's completely dependent on the game itself and what it's trying to do. I attempted to define it in a post above, but no one's commented on it or really explained further outside of "smarter" which is vague

Click to shrink...

Good AI in games is suitably challenging, believable, fair. It nudges players to use their skillset and the game mechanics to their fullest.

People like to define good AI as something that accomplishes its goal (killing the player) as efficiently as possible. But that is never the goal. The goal is to lose to the player while putting up a fight. The answer to the "putting up a fight" part is too often health, damage, hit speed, hit area and so on rather than doing what they do smarter or more efficiently.

 

[ILikeFeet]

Am I underestimating the complexity or is it that devs WANT players to be able to do that if they want to cheese?

Click to shrink...

could be a case of cheesing was unintended and devs didn't really plan for the "pile of bodies in one spot".

 

[Deleted member 15240]

Don't a number of them already incorporate that to some degree? Games like BOTW and RDR2, even as far back as earlier Far Cry games (at least 3, since that was my first), have animals interacting with one another, hunting, etc

Click to shrink...

They do but I think it's usually pretty basic. I want to see things like pack hunting and more complex social dynamics. Animals with separate needs for hunger, hunting, sleeping, social activities, etc that act realistically on those needs. That's more something for park builders and simulation games like Planet Zoo but I think it could work in other games as well. Something like Alien Isolation but instead of a single persistent alien on a space station it's a group of persistent dinosaurs on a small island.

 

[DNAbro]

That's actually kind of an interesting example. That type of thing comes up A LOT in AI discussions. Corpse piles are immersion breaking. But it shouldn't be too hard to write a behavior to recognise that kind of play. X number of corpses in close vicinity + player in hiding = "she's picking us off, flush her out!".

Am I underestimating the complexity or is it that devs WANT players to be able to do that if they want to cheese?

Click to shrink...

It's the second most likely. The fact that it comes up so often means they don't mind you doing it. It's not like it's a rare outlier in any game featuring stealth.

 

[Negotiator117]

https://towardsdatascience.com/artificial-intelligence-in-video-games-3e2566d59c22

Nice piece on game AI covers Alien Isolation.

 

[horkrux]

Strategy AI's in particular should be better enough that when you set it on hard they don't need to give them free resources. That only makes the early game more difficult while the late game becomes incredibly braindead easy.

Click to shrink...

The problem is devs often know their own game a lot less than a dedicated player. So whatever strategies they implement, they are probably really bad to anyone even half-decent at the game lol

 

[Sande]

could be a case of cheesing was unintended and devs didn't really plan for the "pile of bodies in one spot".

Click to shrink...

It's not so much that the AI should recognize a very specific corpse pile scenario, but the general "we're getting absolutely slaughtered, maybe we should do something rather than continue business as usual" situation, which happens all the time. I get silently checking the first body alone, but once anyone knows the whole group is getting picked off, they should alert everyone immediately and enter higher alert.

 

[Kongo B]

Anyone that didn't vote for "I want more complex AI" doesn't have an ounce of imagination.

Better, more complex AI is one of the things that I'm waiting for from this new gen that just started. No amount of 4K ray tracing is gonna do anything for me. I want developers to push in uncharted territories with AI and I want them to push hard.

 

[Bradford]

Good AI can really give an experience more life.

F.E.A.R. is an all-time great due to its AI. TLOU2 on Grounded/Survivor is pretty much the only game I can think of with AI that compares like that. Even Alien Isolation on Nightmare w/ Unpredictable Alien installed doesn't reach that level of tactical prowess.

I really wish more games pushed this.

 

[elenarie]

A simple AI, of course. Generalised self learning AI, if I can have a dream. One which you feed a very basic set of rules per game, and then just let it train via reinforcement learning or something similar.

 

[Boy]

I'd like to see A.i break our gaming routine/strategy.

What i mean by that is when we play a game, i feel like we develop a certain gameplay style to help us get through the game. It would be cool if
the A.i learns your routine, being more aware of what you're going to do ( to a certain extent) forcing you to change up your approach to situations instead of the same old tactics you've used before.

 

[Sande]

Basically this, yeah. It also literally barked shit at random before doing anything and the player would just fill in the gaps between what they heard and what happened.

F.E.A.R. is pretty much a case-study for the fact that you don't actually want smart AIs in games, you just want AI that seems smart.

Click to shrink...

No it's not. It shows that transparency and possibly even deception can give people a more positive impression of AI. And I wish even seeming smart was something more games were interested in.

 

[RogerL]

I think Sports titles have the most to gain here; Madden, FIFA, 2K, etc. advanced sports stats and sports games are about to reach a new level this gen and now that the hardware is there to support more complex systems I have high hopes, I might also just be high though.

I agree that realistic enemies in shooters might take a lot of the fun out of it. It is funny to think game design with dumb AI have in a way formed how we enjoy stealth and horde enemy game modes. More complex and realistic isn't necessarily better.

Click to shrink...

You basically want all players on the field to play like the real player they resemble? If not they are not good enough, thus would need to pass the Turing test for each simulated player... Result would be you playing the worst player, not the hero...

Realistic enemies in shooters would be aim bots - never misses a head shot, you need to work in cover and pop out when they look in a different direction... then they would learn that, and place a hidden sniper...
It could over time force a need for cooperation among the human players - could be interesting.

The other alternative is for games to come out where each human player leads a squad of AI player (or having AI players fill out the numbers) - this would makes battles much bigger!

 

[Shrikey]

It's the second most likely. The fact that it comes up so often means they don't mind you doing it. It's not like it's a rare outlier in any game featuring stealth.

Click to shrink...

Yeah, I'm leaning towards that too. But it's weird to me because Naughty Dog specifically have always been good at driving character movement. In UC they had grenades to keep you moving. In TLOU2 the dogs were a genius way of keeping you moving. And they had enemies drag you from under trucks and whatnot. I guess it might also just be a resource issue. Is it worth the effort in terms of how often it will happen? That said, TLOU2 is like the most over detailed, blink-and-you-miss-it game ever.

 

[Super Rookie]

This seems awesome, but I'm not entirely sure if it'd be ideal for a game dev because it essentially turns the AI into a black box, as opposed to something they can more easily and understandably tweak and alter

Click to shrink...

That's how Forzas Drivatars work I believe but they have variable sliders on them so they don't become insanely impossible but can still give a good challenge but tweaking the parameters.

 

[Dolce]

Good AI is a feeling rather than complexity. You can make AI more complex, but that also means it might make more and more mistakes. This is the actual importance of FEAR: the illusion of complex AI. It's fun to fight enemies because of that illusion, even when they aren't very smart themselves.

 

[808s & Villainy]

That's how Forzas Drivatars work I believe but they have variable sliders on them so they don't become insanely impossible but can still give a good challenge but tweaking the parameters.

Click to shrink...

There are significantly fewer things a car can do than say a NPC human or creature which makes tweaking its AI, even when it's a "black box", much simpler in that regard

For example:

• Recognize the environment and your hitboxes better so you don't just keep hitting walls over and over. Also, have defined strengths and try to maneuver to use them. Like a spear guy is good in narrow corridors. From often places them in corridors, but they're trivial to pull out of them. What if a spear guy not only wanted to stick to corridors but also pushed fights into environments that favor them. Archers would try to get to middle of open areas so they can't be safely approached and so on. Right now enemies almost without exception just beeline your locations regardless of everything.
• Learn from your mistakes. Don't keep using the same attack the same way just to get parried every time or circle strafe backstabbed etc. A good example of this would be that the AI might notice the player always dodges right and uses better attacks or aims the attacks to the right to punish that. At its simplest this could just be a weighting factor on the current randomizer that makes successful attacks more likely than unsuccessful.
• Counter "leashing" which is when players take them to their zone's edge and take pot shots there. Break the leash rules if pushed or just hard return to your post.

In essence I want them to vastly increase the amount of states and possible actions for each state for more believable and varied behavior, and I want events in an encounter to impact those actions for that event.

Click to shrink...

Yeah, but then that becomes a fundamentally different game. The current impositions on the number of actions of enemies, where they show up and where they can roam are all carefully designed to create specific combat encounters with specific difficulties. For instance, if enemies could just follow you wherever you could easily end up in a situation where enemies overcrowd you in an area, or specifically designed "rests" between combat encounters becomes broken, etc.

In terms of your second point, that's also fundamentally different from how these enemies are designed. These action games are fundamentally about pattern recognition. Recognizing an enemy's startup animation, recognizing its range/reach and any quirks in its hitbox, recognizing its recovery time. All these aspects mean that each type of attack involves carefully tweaking of every level of its animation. If an enemy with a spear doesn't have a way to cover for its weakness, that's not a failing of AI or an oversight. It was designed that way to allow users to find that weakness and exploit it. Now yes, there is something to be said about giving enemies enough actions that finding their pattern isn't super simple, but you're never going to get a game like a From game where the enemy can perform arbitrary movements of their weapons, because that's not the game they design and there would be no sense of the player actually improving or getting better at the game if that were the case

I guess my point is that the "good" AI, as you define it (in this situation at least) isn't a matter of devs not being able to, it's a matter of them not wanting to. It breaks the game flow in some cases. It breaks the level design in some cases. It breaks the idea behind specific types of combat and how the games are designed. That isn't to say there isn't room for it. Back to that second point, if an enemy has too few patterns it doesn't become interesting to fight either. So there's definitely a balance between too few, and so many that there's no way to learn their patterns.

 

[Super Rookie]

There will definitely need to be a balancing act with frustrating AI and smart but fun AI too.

Just imaging what a "Director" AI could be like in a new Left 4 Dead game that learns from every game...

 

[madgorillaz]

I don't think AI needs to be "harder", but more complex. Specifically, I have three big things in mind:
1) Becoming more aware of the environment
2) Dynamically reacting to the player's short- and long-term playstyles
3) Greater interaction with other AI behaviors/states

Note that TLOU PII took huge strides with #1 and #3.

I think this will be an obvious area for growth not for the PS5 generation, but for the PS6 generation, where I expect machine learning to inhabit a greater role in the hardware design.

 

[Chief Malik]

Companion AIs are so dumb, as is NPCs. Please fix this.

 

[Gestault]

Most people don't want "better" AI. Traditionally fun enemies, ironically enough, broadcast their current and upcoming behaviors and have a pretty limited set of possible actions. The best AI is a fit for its gameplay sandbox, and is relatively transparent about what it's doing from the player's point of view. 99% of the time, any limitations/shortcomings are related to engineering and game design, not technology. Heck, a lot of times, successful game "AI" is as much about superficial aspects of enemy designs as it is actual behavior. And unfortunately, most machine learning (at the micro-scale) would only result in more effective AI counterparts, which would make a pretty miserable experience if left on its own.

 

[kurt]

AI is what the difficult should be in a game. Most hard games are based on low healt or insane projecttiles or bad/slow gameplay (aka dark) :/

 

[kitler53]

I agree that if developers made their AI too good at the game it wouldn't be fun in some cases, but rather than making it more "skilled" they should instead just aim to add more variance, so that its reactions don't become predictable and thus exploitable

Click to shrink...

i'm kind of agreeing and kind of disagreeing here. i think you are on point with people exploiting games when they figure out a way to easily win. but also many people are just monontous in the way they play.

what i have in mind is an AI who's focus is to learn how the player is playing and adjust their strategy to exploit them. thus the only way to beat the AI is to be constantly evolving my strategy and use all of the gameplay mechanical available to me. i am soo guilty of falling into a "1-trick pony" kind of gameplay loop when it's working.


Also, i wish more games would apply AI and science to difficulty. there has got to be a science to how many times per hour i'm willing to die in game and still be having fun. 60 times per hour is no fun. 0 times per hour is no fun. for me maybe the best number is 5 and the difficulty could be adjusted in real time to target the level of difficulty that perfectly suites my skill level.

 

[Old Gimlet Eyes]

AI garbage state across all genres is what prevents me from playing single player games. I much, much prefer the random unpredictability of multiplayer matches just because you never know what you will get.

 

[jotun?]

It depends on the genre

Fighting games could stand to have more complex, natural-feeling AI. Same for RTS and basically anything where the AI is just a stand-in for human opponents.

Stealth games are probably the most difficult to balance. You want NPCs to be simple enough to have consistent and obvious ways to manipulate them, but you also want them to appear smart enough to maintain immersion, and I don't know if there's a way to fully accomplish both of those. Many games have to make this balance, but stealth is the most extreme example I can think of.

 

[Super Rookie]

Better AI for NPCs where they wake up and decide what to do for the day ahead would also be amazing instead of scripted routines.
Shenmue 4...

 

[Poimandres]

I want games to expand upon and explore what STALKER was doing with AI. Having NPCs that roam the whole map and interact with each other independent of the player makes a game seem so much more alive, and adds greatly to replayability.

Seriously, load a save in STALKER and the next 5 minutes could play out completely differently. Maybe you come across a wandering group of bandits, maybe you stand back and watch the bandits in a battle with a pack of dogs, maybe the bandits are all dead and you loot their corpses.

 

[SuzanoSho]

I want AI's that come up with interesting new strategies.

Also I want ML AI's becaus just watching them play against eachother is lots of fun!



Like... look at them finding game exploits!

Click to shrink...

Holy speedrun AI!...

 

[Sande]

There are significantly fewer things a car can do than say a NPC human or creature which makes tweaking its AI, even when it's a "black box", much simpler in that regard



Yeah, but then that becomes a fundamentally different game. The current impositions on the number of actions of enemies, where they show up and where they can roam are all carefully designed to create specific combat encounters with specific difficulties. For instance, if enemies could just follow you wherever you could easily end up in a situation where enemies overcrowd you in an area, or specifically designed "rests" between combat encounters becomes broken, etc.

In terms of your second point, that's also fundamentally different from how these enemies are designed. These action games are fundamentally about pattern recognition. Recognizing an enemy's startup animation, recognizing its range/reach and any quirks in its hitbox, recognizing its recovery time. All these aspects mean that each type of attack involves carefully tweaking of every level of its animation. If an enemy with a spear doesn't have a way to cover for its weakness, that's not a failing of AI or an oversight. It was designed that way to allow users to find that weakness and exploit it. Now yes, there is something to be said about giving enemies enough actions that finding their pattern isn't super simple, but you're never going to get a game like a From game where the enemy can perform arbitrary movements of their weapons, because that's not the game they design and there would be no sense of the player actually improving or getting better at the game if that were the case

I guess my point is that the "good" AI, as you define it (in this situation at least) isn't a matter of devs not being able to, it's a matter of them not wanting to. It breaks the game flow in some cases. It breaks the level design in some cases. It breaks the idea behind specific types of combat and how the games are designed. That isn't to say there isn't room for it. Back to that second point, if an enemy has too few patterns it doesn't become interesting to fight either. So there's definitely a balance between too few, and so many that there's no way to learn their patterns.

Click to shrink...

I'm not suggesting anything that breaks the fundamentals of Souls games though. All of those just reduce the monotony and abusability (and not the good kind imo) and promote more creative solutions.

Like the leashing example, it would only happen when the player has clearly abused the leash mechanic. You wouldn't accidentally overwhelm yourself with dozens of enemies. And none of the examples ruin the pattern recognition aspect. I'm not suggesting changing combos or wind-up times or anything like that. Though I think there's potential there too but at that point you actually would be taking a pretty fundamental step away from Souls.

 

[FTF]

I mean who doesn't want better/more complex AI?

 

[Sande]

I mean who doesn't want better/more complex AI?

Click to shrink...

"You don't want better AI. You think you do, but you don't." is one of the most common answers every time this topic is brought up.

I think it's absolute horseshit that comes from the incorrect idea of what good game AI means that I briefly mentioned in the OP.

 

[HeyNay]

When people say they want better AI they're not saying "I want enemies that can outsmart me each time", or "I want a more difficult game". They're saying they want more complex reactions and deeper levels of interactivity with opponents. Behavior can still be tailored for the optimal difficulty level while also behaving in new and unexpected ways. Most of all, I think players want to be surprised by how characters react to them on the fly. A lot of games that feel fresh and exciting in the beginning lose their novelty once you know what to expect, and that is where AI fails and the illusion breaks down in the vast majority of games.

 

[LCGeek]

Others have highlighted the real issue how the gets used may not benefit the game. Some consumers don't really want a mega hard challenge based on AI, they wouldn't mind if that power was used to design better dynamics or allow for more of them. DLSS is proof AI can be wielded in other non typical ways.

Really on devs to to do something bold and implement it well.

It's a nice thought but it's frought with practical problems when rubber meets the road. Should be done but it's quite the gamble of investment money which games are short on.

 

[Nooblet]

It's the second most likely. The fact that it comes up so often means they don't mind you doing it. It's not like it's a rare outlier in any game featuring stealth.

Click to shrink...

It's probably down to the fact that you can't move bodies in TLOU2 and you can't really choose where the enemy goes and stands when they come to "investigate". So if you kill someone then they stay at or around the spot where they came in to look for the body (since you can move a little bit while doing the kill animation), as such giving players a pass there by letting them build a corpse pile is the better option there as otherwise the player will just get frustrated at being detected by something that they have no control over.

 

[LCGeek]

I am not memeing. For FPS games AI just has not gotten much better since that era. In other games? 100%

Honestly I think the next step is Machine Learning AI, that opens up a lot of interesting doors!

Click to shrink...

In quite a few games I still play fps games that lack the depth og halo, perfect dark, and og UT in options offered to players or depth.

Fear is a solid place to mention a bulk of devs haven't really evolved big time over.

Yes we have had improvements have they been worlds apart, that's the debate.

 

[Weltall Zero]

Getting better AI is something people often ask for, particularly when talking about next gen features. This discussion is almost inevitably shut down with "good AI is not fun, actually" or something vague about machine learning.

First, I'll leave machine learning to people who know more than I do and focus on incremental improvements on current AI.

Good AI = bad is a misconception that stems from the idea that if devs do everything in their power to make enemies as difficult as possible, the game will be miserable. Of course it would be, but no one is asking for that. An example of this would be a stealth AI that recognizes a crucial objective, flocks to it en masse and everyone is constantly checking each other's backs, spinning around wildly etc so there's no way to approach the encounter with stealth. Similarly once in combat, everyone would have perfect aim and split-second reflexes so the player gets instantly shredded. Good AI can be bad when approached like this. That does not mean that good AI is bad.

I would say that bad AI = bad. Enemies that fall for the same moves over and over with no capacity to act differently based on history or circumstances are mind-numbing. The same goes for enemies that act exactly the same way every time with maybe a few random variations. An example of this would be an enemy going "guess it was the wind" and going about their day when they see their friend's headless corpse or seeing a 3 feet tall corpse pile and deciding to investigate it with no backup.

What I see in examples like this is not the pinnacle of how fun AI can be. I see something that takes me out of the experience and keeps everything samey even when there's a clear opportunity for more emergent and unique scenarios. Players want to be challenged. Expanding and tuning your AI's toolbox can add to that in much better ways than inflated health pools.

Click to shrink...

Not to play a credentials pissing contest, but having a background in both game development and AI, seeing the thread open with a complaint about the conversation being "shut down" with "good AI = not fun" and "something vague" about machine learning doesn't exactly inspire a lot of confidence that you're too terribly interested in the theory and concepts involved in either, versus plain old venting off and telling devs how they're doing it so obviously wrong.

 

[Sande]

Not to play a credentials pissing contest, but having a background in both game development and AI, seeing the thread open with a complaint about the conversation being "shut down" with "good AI = not fun" and "something vague" about machine learning doesn't exactly inspire a lot of confidence that you're too terribly interested in the theory and concepts involved in either, versus plain old venting off and telling devs how they're doing it so obviously wrong.

Click to shrink...

The language I used was definitely not meant to signal I'm not interested in game design. I wouldn't have made the topic if I didn't find the subject important and fascinating.

I have a passing familiarity with machine learning, but like I said I'm not interested in making this topic about that. Others are free to discuss it, of course. I know how traditional AI works quite a bit better, but have no hands-on experience beyond simple hobby projects. But I think everyone can discuss a topic like this generally, without programming specifics or dev background. After all, the main argument here is whether players would like any kinds of improvements in AI, not the nuts and bolts implementation on dev side.

Edit: And just to clarify, the opening paragraph was referring to the parroted forum nonsense that always crops up. There is a legitimate discussion to be had about the pitfalls of treating more advanced AI as having intrinsic value to the player experience. But that discussion has been mangled through so many steps that it has boiled down to "good AI = bad" by the time it lands in Era. I'm trying to argue that hardline mantra, because basically no discussion can be had until people move past it.

 

[Tokyo_Funk]

Personally, in single player, I want them to be smart enough that they're believable, but not smart enough that they're undefeatable. It would be nice to see more emotional states such as fear/withdrawing, search patterns that felt more realistic and such.

In MP games though, the more close to people the better. Times when I don't have net access it would be good to have bots that actually work together as a team on both sides.

Kind of a shame that this did not go further:

 

[Chaos2Frozen]

What people tend to want is the illusion of a 'fair' and 'smart' AI and being given a chance to outsmart them and feel good about it.

Mark Brown has a video on AI design a while ago

 

[TheMadTitan]

Nothing has topped MGS2's.

People like to overlook how enemies. Meanwhile I see people bringing up MGS5 when the entire area will be alerted once someone screams in the middle of a sandstorm

Click to shrink...

Exactly what I was thinking. What MGS5 does right though is that some point. the game doesn't drop back into the neutral state and everyone remains on the evasion level. Take MGS2 and don't reset the world state because you haven't been spotted yet and that'd be the perfect base routine for action game enemies.

 

[Jawmuncher]

It's fine. They get too smart and the gameplay isn't as fun.

 

[Cyclonesweep]

That's actually kind of an interesting example. That type of thing comes up A LOT in AI discussions. Corpse piles are immersion breaking. But it shouldn't be too hard to write a behavior to recognise that kind of play. X number of corpses in close vicinity + player in hiding = "she's picking us off, flush her out!".

Am I underestimating the complexity or is it that devs WANT players to be able to do that if they want to cheese?

Click to shrink...

The issue with too smart of AI is the game becomes outrageously hard as most people will constantly being outsmarted by the AI. It's a really fine balance.

 

[Hazzuh]

An example of perfect video game AI, goombas in SMB1:

How exactly would the game be improved by giving them more complex behaviour?

 

[Weltall Zero]

The language I used was definitely not meant to signal I'm not interested in game design. I wouldn't have made the topic if I didn't find the subject important and fascinating.

I have a passing familiarity with machine learning, but like I said I'm not interested in making this topic about that. Others are free to discuss it, of course. I know how traditional AI works quite a bit better, but have no hands-on experience beyond simple hobby projects. But I think everyone can discuss a topic like this generally, without programming specifics or dev background. After all, the main argument here is whether players would like any kinds of improvements in AI, not the nuts and bolts implementation on dev side.

Click to shrink...

Don't worry, machine learning (or pretty much any kind of AI that's not rule-based) isn't particularly relevant to games anyway. It's hard enough to debug AI when you know exactly what's supposed to do.

I think if you want to tackle the topic seriously, you first have to define what "good" AI is. It may seem intuitive enough, but that intuition does not only not translate well to any kind of formal definition, it can be easily interpreted in ways that are diametrically opposed.

I would start with at least acknowledging that there's several possible definitions of what constitutes "good AI":
a) AI that straight out wins the most often (you could probably subdivide this into a1, a2, etc. depending on who it wins against, e.g. other humans of different skill levels, other AIs, etc.).
b) AI that emulates the behaviour of an actual human the closest.
c) AI that feels the most fun to fight against.

With these definitions in mind, I think the point you're trying to make is that an increase in a) could translate into an increase in c), either directly or through an increase in b). I personally find that b -> c is a much easier connection to make than a -> b or a -> c, and in fact have experienced first hand how strongly a and b are at odds with each other. To put it bluntly, I've spent almost as much time toning down my own AI to have human reaction times, make bad calls, etc. as I've spent coding the AI in the first place (and AI that was pretty dumb to begin with), and the difference in enjoyability is astounding. So yes, I very much subscribe to the notion that limiting even the simplest AI can be more important than giving it particularly smart decision-making capabilities.

 

[Dyno]

Personally I just want more expansive AI. More variety of actions and behaviours and more interactivity between them and other AI as well as the player.

 

[Jaychrome91]

Killzone 2 had great AI imo.

 

[EvilBoris]

This almost seems like a meme at this point. It seems INCREDIBLY reductive to the numerous AI systems that have since, objectively, topped FEAR. We know what FEAR did, AI wise, and most AI systems nowadays are far more complex and far more reactive.

Click to shrink...

haha yeah , every single thread about Ai.
Elites roll when you throw a grenade!!
Grunts run away when you kill an elite.
Nothing has beat F.E.A.R!!

 

[Sande]

Don't worry, machine learning (or pretty much any kind of AI that's not rule-based) isn't particularly relevant to games anyway. It's hard enough to debug AI when you know exactly what's supposed to do.

I think if you want to tackle the topic seriously, you first have to define what "good" AI is. It may seem intuitive enough, but that intuition does not only not translate well to any kind of formal definition, it can be easily interpreted in ways that are diametrically opposed.

I would start with at least acknowledging that there's several possible definitions of what constitutes "good AI":
a) AI that straight out wins the most often (you could probably subdivide this into a1, a2, etc. depending on who it wins against, e.g. other humans of different skill levels, other AIs, etc.).
b) AI that emulates the behaviour of an actual human the closest.
c) AI that feels the most fun to fight against.

With these definitions in mind, I think the point you're trying to make is that an increase in a) could translate into an increase in c), either directly or through an increase in b). I personally find that b -> c is a much easier connection to make than a -> b or a -> c, and in fact have experienced first hand how strongly a and b are at odds with each other. To put it bluntly, I've spent almost as much time toning down my own AI to have human reaction times, make bad calls, etc. as I've spent coding the AI in the first place (and AI that was pretty dumb to begin with), and the difference in enjoyability is astounding. So yes, I very much subscribe to the notion that limiting even the simplest AI can be more important than giving it particularly smart decision-making capabilities.

Click to shrink...

The definition is certainly tough to nail down which is a pretty big reason why these discussions can go in circles.

With your categories my central point would be: Don't assume that good AI only means option a. And don't assume that option a has to be taken to the most extreme point possible at the cost of everything else.

 

[masterceewhy]

We sorely need better AI, I really hope it gets better within at least the next year or two.
I feel like I'm going up against the same opponents regardless of the game because the AI is so bad.

Like can you guys really feel a difference between COD enemies to Destiny enemies?
Or Watch Dogs stealth to any open world game with stealth?

AI is probably the worst part of games and it needs a big overhaul this generation, regardless of the hit to graphics.

 

[HeWhoWalks]

My top four games (in this order)...

- F.E.A.R.
- Killzone 2
- The Last Of Us Part II
- S.T.A.L.K.E.R. Shadow of Chernobyl (I know console-only gamers won't understand, thus, it is often left out of discussions like this).

Outside of those, most games leave much to be desired in this area.