Talking about game AI

We have all seen the AI bait-and-switch with different titles. Is this a marketing decision to make the AI less AI and more AD (artificial dumbness)? They don't want the game to be too challenging because then people get frustrated and don't play it. So they dumb it all down to the lowest common denominator. Making it easier to play.

Or is it a technical hurdle that requires more work on the game and they release with a simpler AI to reduce the amount of development and QA needed and therefore release the game sooner and with fewer bugs/glitches?

I am speculating, since I don't know, that the decision is due to technical limitations because otherwise they would leave in the option for the best AI. That way the player chooses the AI level.

I got bored recently and played Uncharted 3 on higher difficulty and it really was barely more difficult and it didn't seem to have anything at all to do with the AI. It was more about losing health quicker. Maybe less loot too, but sure.

One of the most common problem is that human remember things they think about more. An AI that does something stupid is something we remember better than a smart action AI may have done.

I think AI may not matter as much as people look at. For shooter good staging and scripting of the enemies makes they game better than AI.

Second what are using to compare the current AI to?

AI is a complicated topic, and the idea of true AI is something that's barely understood even in the world of acedemia. When people say they want "good AI in their games" what they really mean is they want enemies to show up, provide some challenge, and die in an interesting way. It doesn't really matter if the AI character is making dozens of intelligent decisions about what action/path/whatever to take next because that's not necessarily reflected in behavior. It is a true statement, though, that the more streamlined the AI of a character the less likely that the character will do something screwey.

However, deadly enemies doesn't necessarily mean the AI is smarter, either. Yes, this can factor into the tactics of NPCs but part of programming AI is making the enemies that don't behave like computer programs. There's no reason why every enemy in every game can't shoot you with 100% accuracy all the time (since they are basically aimbots) but obviously that's not fun. It's a very difficult balancing act to make them seem like imperfect but thinking beings.

-Byshop

All good points. It is complicated. But in this case the definition matters. I think your definition in this casual context is correct.

But is it a marketing or technical decision to reduce the AI (still using your definition) to the point of idiocy? Seems like a marketing decision to me. But it is a delicate balancing act as you said. Maybe the solution is to allow the player to determine which way they want things to work. Something more comprehensive than just Normal or Hard or Insane, etc. Perhaps they could design an AI configurator like they do for the character hair, face,eyes, etc. But instead you decide an alert radius let's say. Inside the radius NPCs react to noise/movements. Then you could choose what the NPC does when alerted. Move away and hide, go stealthy, come at you, call in back up, etc.

@Byshop: You're right about there being a difference between difficulty and AI. The point of games is to be fun and entertaining, not to simulate real intelligence. But part of what makes enemies fun to fight is the illusion that they are smart, even if they really aren't. And sometimes that has nothing to do with game difficulty. You said that "it doesn't really matter if the AI character is making dozens of intelligent decisions about what action/path/whatever to take next because that's not necessarily reflected in behavior." I would say that intelligence or at least the appearance of intelligence should be reflected in an enemy's behavior, even if its all smoke and mirrors.

Its marvelously entertaining to sneak up on NPCS in games to observe them when they aren't trying to kill you. Do they just stand in place waiting for you to trigger their attack mode? Or do they talk to each other and explore the environment? In stealth games, do the enemies just follow the same walking path over and over again? Or do they improvise sometimes and say different things?

A well designed game creates the illusion that these characters have personalities and lives beyond the moment you fight them. So no its not about having true AI, its about making the enemies interesting and entertaining. The illusion of intelligence usually goes hand in hand with that.

It's not a question of "reducing the AI", it's a question of trying to do something and not succeeding. Even in a realtively simple action game you are trying to program behavioral routines for literally infinite combinations of sitautions. You can talk about actions like "come at you", "go stealth", "move away and hide" but that's ignoring the massive amount of complexity that goes into making any of those actually work.

-Byshop

Okay, you're saying it is a technical issue and it's too complicated. From my naive point of view it doesn't seem so complicated that it couldn't be done. But it is probably not been done because it is hard, takes time and lots of money. The benefit of having better AI may not translate to better profits. So hard to justify the cost.

I think that better AI means a better game and a more memorable game. They stand apart from the rest of the pack.

@calicobass: it takes a lot to program great A.I. in games, and even if you have great systems in place to prevent typical A.I. idiocy, there's no such thing as a perfect system. It might seem simple, but in reality it's a big struggle to make great A.I. in games. Programming bugs pop up, unintended behaviors show themselves, and rules can overlap and cause problems.

Man, I'm with you on this, but are you aware of the fact that the most popular enemy type these days are fucking zombies? Do you think modern gamers care about complex AI routines? They are perfectly fine with enemies that run at them in a straight line.

Now, those of us who demand more can always go back to F.E.A.R. and SWAT 4 (those lucky enough to own a copy of the latter) for our good AI fix.

It's staggeringly complex. You can spend a ton of money of money trying to make the most complex AI, but the end result behavior might not -look- that different from a game that doesn't put nearly as much effort into it. A great example is the game Jurassic Park: Trespasser. It was probably the very first game to use a realistic physics engine as a part of its gameplay model. Dino AI was also supposed to be pretty top notch. They would fight each other and make intelligent decisions, but in the end for all that effort the game ended up just not being that good so none of it mattered. The dinos may have been smarter than the AI in other games but in terms of actual gameplay their behavior didn't really seem all that different from "dumber" games.

-Byshop

I wish I knew more about how the "AI" gets implemented in games but I don't. I think we agree it's a hard problem to solve but that means it's ripe for the picking for someone to capitalize on a solution that's easy to reuse and works reliably.

Maybe the current approaches are coming at it from the wrong angle???

@calicobass: there's only so much you can do with programming, honestly. Especially making it so the A.I. doesn't become frustrating so players actually enjoy playing

Exactly. As I mentioned earlier, a computer programmed character need never miss so good AI programming is just as much about believably fallible while not making them incompetent.

AI as a concept is so complex and difficult to quantify. There's so much that we take for granted that our brains do for us behind the scenes to allow us to function. Language is a good example. Hearing words to understand what someone is saying is only a small portion of real communication. When you talk to another human being, the conversation takes on a context and based on that context you generally will have an idea of what the other person is going to say before they say it and that's a huge part of understanding. If you're talking to a friend about a breakup they just had, you have the context of understanding relationships, the emotions involved, the fallout of when they end, as well as basic human needs. If someone in the middle of that conversation says something that makes no sense for the context like "throat wobbler mangrove!", our common response isn't "why did you say throad wobbler mangrove?". The common response is "wait, what?!?!" because the listener would in most cases assume they misheard because what the other person said doesn't make any sense to them. A computer, by comparison, has no idea of context. This is why pretty much every voice to text program doesn't work very well. Computers have to rely on hearing each word to try to translate them into known words in the English language, but people often don't pronounce every word clearly. Words run together and they get mush mouthed, but people understand what they are saying because they get the context of what is being discussed. A computer can't do this.

So yeah, AI is hard. In reality what we mean by AI in computer gaming isn't really independent thought but pre-programmed behavior routines. Take "The Last of Us" for example. I'd use that as both an example of good and bad AI programming. Infected enemies follow patrol patterns, but if you are spotted or they hear a noise they go into an "aggitated" state that they stay in pertty much forever. You have to factor this in before you throw that brick to distract an enemy because while it might solve your short term issue if they being in your way, it might make your job harder in the long term.

However, TLoU also has one glaringly obvious AI issue and that is that your companions are invisible to enemies. This was obviously a technical scope decision that they had to cut, because getting your NPCs to succesfully stay out of sight and not accidentally give away your position was probably extremely difficult to do. Good pathfinding (especially "stealth" pathfinding) is extremely hard to program.

Now look at MGSV. This game is known for some pretty good AI. When you make your presense known, enemies will start attacking your position. They don't tend to do stupid things like take cover that's perpendicular to your position making themselves easy to be shot. Instead they take smart cover, they will start shelling your position if they have mortars, they launch flares into the sky to light up the area if it's at night, if they lose track of you they start sweeping the area starting with your last known position, etc. Also, if you score a lot of headshots then they start wearing helmets. If you attack at night a lot they start heading NVGs. However, these are really more like programmed behaviors. They aren't decisions being made by the computer, they are decisions that were made by a human ahead of time and are set to go into effect when certain conditions are met. This is not so much AI, but a lot of scripting put into the game and there's a difference. This kind of scripting works better in certain types of games but not as well in others.

So yeah, it's complex. If you'd like to know more about it, here are some books on the topics. Some of which are written by the programmers for some of teh games that I've mentioned:

http://www.amazon.com/s/ref=nb_sb_noss_2?url=search-alias%3Dstripbooks&field-keywords=programming+video+game+ai

-Byshop

Until we're able to design programs or chips with neural networks that would be capable of learning on their own (If such a thing is possible), programmed AI is probably going to remain passable at best, and even if such a breakthrough happened, the results may not be all that enjoyable. Also, it would immediately present a laundry list of moral and ethical implications, as who could say if such an AI would lead to what could technically considered consciousness at some point?

But hey, I'm just going by what James Cameron taught me when I was a kid, what the hell do I know? It's still fun to speculate. But I don't see any reason why synthetic neural networks wouldn't be theoretically possible when technology gets to the point that could afford it, and quantum computers are on the horizon which will exponentially increase processing power far past what our best supercomputers can currently do.