Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Arx]

More specifically, the Starliner lost track of time and some piece of control software thought it was somewhere it wasn't. As a result, it had to burn more fuel than planned (not sure if that's because the bug burned fuel, more fuel had to be burned to correct the issue, or both) and couldn't get into a stable orbit to rendezvous with the ISS.

[Madman198237]

Sounds like it started its insertion burn early. If you play KSP, you might know that the ideal place to circularize your orbit (i.e., raise periapsis/perigee to a point outside the atmosphere) is at your apoapsis/apogee. Burning earlier is inefficient, inefficiency costs fuel, fuel is in short supply in the high-precision low-margins world of rocket science. Add in a wrong burn and even if you *did* thrust in the correct direction (it may very well have been pointing in the wrong direction for its burn) you'd very easily run out of fuel before reaching intended orbit, or, as actually happened, reached the intended orbit without sufficient fuel to make the later maneuvers to rendezvous with the Station.

[Trekkin]

More specifically, the Starliner lost track of time and some piece of control software thought it was somewhere it wasn't. As a result, it had to burn more fuel than planned (not sure if that's because the bug burned fuel, more fuel had to be burned to correct the issue, or both) and couldn't get into a stable orbit to rendezvous with the ISS.

From what I understand, this is fundamentally correct: the main engine got an incorrect MET corresponding to after the needed burn, but the RCS got the correct MET and so was burning fuel keeping the nose pointed the right way during a burn that didn't happen.

[Reelya]

I've decided that from now on whenever I hear "AI" I substitute the phrase "advanced pattern recognition and extrapolation."  I think if they'd have called it that from the beginning people would not think it is so magic. After all, that's basically what every type of "AI" we have today does.

It also avoids the annoying debates about "is it intelligent" or ethics or whatever.  A "machine that has advanced pattern recognition and extrapolation capabilities" sounds way less existentially threatening than "an artificial intelligence".

(my bolding)

But even that use is context sensitive.  Just about no game AI even aspires to such lofty heights...

There's a reason for that. It's not a lack of processing power or a lack of good algorithms, it's that the layman vs the game designer have a totally different idea of what the purpose of "game AI" even is. Game AI isn't put in with the goal of "optimal decision making to beat the stuffing out of the player", because that's not actually something that produces games that most people find fun, to the point they'll pay good money for it.

Game AI is about supporting the single-player experience and pacing they're going for. If you suck, it has to suck, if you git gud, it has to git gud. That's why you get RTS stuff like units that sit in a specific location, only attack in a certain range, and when you move away, they move back to their original spot. It works for the experience they're going for, and it works about as well for both sucky and good players.

[wierd]

There's something to be said about "Nintendo Hard" though.  Punishing game design, including "good AI" that does not cheat, but makes good and solid decisions based on the same kinds of data the player would get, would definitely appeal to some players.

The "Filthy casuals" are the ones that would balk.  Better would be to give an option in the options menu about that.

[feelotraveller]

Sure.  But that is a different discussion.

The initial topic is about why the hell one would call that function/process an AI when it is clearly not ever intended to be one.

[Reelya]

There's something to be said about "Nintendo Hard" though.  Punishing game design, including "good AI" that does not cheat, but makes good and solid decisions based on the same kinds of data the player would get, would definitely appeal to some players.

The "Filthy casuals" are the ones that would balk.  Better would be to give an option in the options menu about that.

Nah, that comes under the rubrik of "things the player can't even see".

A lot of the time they put effort into elaborate "properly simulated" stuff, but it ends up being a complete flop, since the inner workings of some AI or simulation aren't visible elements. It ends up being a lot like movie logic: the superficial appearance of "being right" is more important than actually getting it right. That's why you constantly get big lists of how movie logic "gets it wrong". They're already aware it's wrong, it's not a mistake, it's deliberate, because they know that it works cinematically, which leads to the thread-relevant bit where in every movie some viewer always says "but that's not how the science even works, why can't they get it right?"

That's also the reason that "not technically correct" simulation code continues to be used in games.  Games work when "players does action X leading to enemy response Y" is true. But, this is not how the real-world even works. Responses in the real-world to your actions may be unpredictable and fickle. "Good" game-AI would have a ton of hidden variables and it wouldn't have that satisfying linear response to the player's actions. Arguing "Nintendo Hard" isn't the same as arguing for punishing realistic AI. "Nintendo Hard" is about memorizing the patterns, and if the AI can change the patterns those hard-fought "Nintendo Hard" skills are worthless. Imagine in a "Nintendo Hard" game if you killed 6 grunts using one attack pattern, then the 7th guard was smart enough to notice how you killed the other guys, so he changed his pattern and killed you instead. Do you expect the "Nintendo Hard" fans to be all over that? Additionally, realistic AI would be almost impossible to "tune" easily for difficulty levels. It's not a linear quantity where you just "add more AI" and the game gets incrementally harder in some predictable fashion. You'd have to train the AI independently for all desired difficulty levels.

A really smart AI would trick the player, hiding their actions, then surprising the player and beat the stuffing out of them, and you wouldn't know why it happened. This is the optimal "real AI" version of any game with any sort of Fog-of-War. It's what a properly written AI would do, but players wouldn't know the difference between that and cheating. There are quite a few stories of developers building "proper AI" for RTS, getting accused of the AI cheating, then they yanked the AI and put in a dumb one (using the standard production bonuses to set difficulty level), and the accusations stopped. What actually matters in game design, both for the Nintendo Hard crowd and the casuals is that you get predictable feedback based on your own actions. Realistic AI undermines this assumption. also, realistic AI only conceivably works in games which are exactly symmetrical. Any sort of asymmetry and it going to destroy the game.

[ChairmanPoo]

There are quite a few stories of developers building "proper AI" for RTS, getting accused of the AI cheating, then they yanked the AI and put in a dumb one (using the standard production bonuses to set difficulty level), and the accusations stopped.

I'd like to see specific examples of this because in my experience it's more common for developers to make fake claims about their game's AI.

[Reelya]

I think it came up in a video by the guy who coded the AI for Civ 3 and Civ 4, was also mentioned in a video "the rise and fall of Empire Earth", how they accused one of the sequels of cheating for being too good so it was then nerfed. It's partly down to that people expect the AI to cheat, so if things feel "unfair" then the "cheating AI" accusation comes out. The point is, it's not something you can see. The cheating accusation flies even when it's just the computer enemy making a good move. Fog of War allows players to rationalize their losses this way.

This is from devs: they've basically found that if they try and program something really smart that the accusations of cheating fly around anyway, so realistic AI been put in the "why bother" bucket: costs too much, takes too much time, is unreliable and hard to tune, and people will just accuse you of being assholes even if it works as intended. It's a project-killer.

There's also a very fine line between being just being superhumanly good and cheating. For example, Aim-bots are clearly cheating if used by a human player, but what about playing an actual bot? Is that also cheating if they have perfect aim like an aim-bot?

[ChairmanPoo]

Well no. There are ways to find whether the computer is cheating (not least map hacks). I mever played EE for any lemght because it was boring, but for instance, RA2, despite the dev's claims about the AI, was an obvious cheat

[wierd]

I agree.

In my experience, the cheating of the AI, is more that it ignores resource restrictions, or does not obey cooldowns correctly, and knows where everything is on the map, and thus plans moves with information it should not have.


I can test for this, and prove it, by placing the AI into a custom map where I have purposefully restricted the resources available to a hard limit, and seeing what it produces, how fast, and where it places them.  This is especially true if there is an observer mode, and I can enlist the help of another person to just sit there and WATCH the AI, and time its action speed.  This is typically how I verify cooldown compliance.

I have done this a few times, and shown that yes- the AI is a cheating mo-fo.  I have even witnessed the AI simply CONJURE units out of thin air, just outside the player's visible circle, and attack with them.  Seriously-- that's fucked up shit.



When I find that yes, the AI is indeed a cheating bastard, I get very hostile about it.  To me that's not a skill challenge, that's seeing how well you can interject on its behaviors, and anticipate its behaviors to do that interjection. 


 Being able to trick the AI into thinking you are someplace you really arent with decoys is impossible when it cheats, for instance.  It removes whole areas of potential play, because the AI is written for whiny bitches who 1) dont want to invest the time in a good AI, and 2) whine when the AI is good-- where appropriate to lay those condemnations.

Rather than just roll over and go "Oh right, sorry-- we will turn on baby mode for you." the devs should go "So, you did structured testing to show that our AI is cheating?  Where-- we spent a great deal of time to make it not cheat. We are interested in it not cheating."  That moves the ball back into the whiner's court, since now they have to back up their butt-hurt with a test result.

[ChairmanPoo]

Not to mention that scaling difficulty with a really good AI wouldnt be that difficult.... just ramp down their available resources

[Reelya]

My main point is that people say they want "proper AI", but if you gave it to them, almost nobody would buy it. Not enough to fund development at least. The most likely result of trying to shoe-horn "good AI" into a game would be something that's only marginably playable. There isn't actually any proven benefit to gameplay other than being able to wank on about how the AI is realistic.

Also my point wasn't whether AI is cheating or not, is that not having the cheating AI doesn't actually prevent accusations of having cheating AI. Sure, there are "ways to detect whether the AI is really cheating" but the vast majority of people making any sort of internet accusations haven't actually done any checks.

Not to mention that scaling difficulty with a really good AI wouldnt be that difficult.... just ramp down their available resources

Th

Not to mention that scaling difficulty with a really good AI wouldnt be that difficult.... just ramp down their available resources

That's just not correct. Complex simulation code is very rife with non-linear problems. You're likely to get unexpected shit happening. It's not like people haven't tried this, it's that it's doomed to failure. Game developers need to ship product, and they need to ship it in a moderate time-frame, it needs to ship tuned and balanced as best as possible. Putting expermental, random, complex and unpredictable code such as solid AI, which is usually NNs of some sort is a fool's errand, if you're expected to pay employees. If you're a single indie dev with no employees and preferably no spouse/children to pay for, go for it then. Meanwhile, if you're someone making something and other people's livelihoods and families actually depend on the product succeeding, steer clear of "magic bullet" AI "solutions" to the problem of building a game.

In the real world, real-life experience is that if someone leading your dev team has some wild idea for some cool AI that's going to drive everything, it almost always fails and goes to shit. Pretty much every one of those examples of game AI that didn't live up to expectations was because someone had your idea about using "real" AI and went with it, and the shitty shit that doesn't live up to the "hype" is exactly what you get when you try that. The promise of "real AI" is that it's going to be this magic sauce that makes everything clever and fun and balanced, and you won't have to hand-craft or kludge anything. If such a thing ever worked, they'd all be doing it rather than kludges. Nobody wants to hand-balance a complex game and add a zillion kludges to make it work. It's very time-consuming.

They always end up either stripping that whole system out and replace it with some extremely kludgy code that is worse than what they would have used if they'd just hand-scripted everything from the start, or they pile kludges on top of the "magic AI" that was supposed to solve all the problems, and you get a system people will be complaining for for years to come.

[wierd]

I would be happy with a simple algorithm if it simply had properly restricted inputs that better represent what is available to a player, and not an omniscient computer-- and obeys actual game rules. (Resource costs, build times, and cooldowns.)


The actual HOW it decides where to move, as long as it is not "Move mass of units to player base location" in nature-- and instead, "Triangulate base location from vectors of enemy unit travel as they enter my visual space"-- is moot.  Use whatever algorithm you want. It need not be NN or a GAN, or anything fancy like that.

Just don't fucking cheat, and dont pretend that the game is even remotely about skill if you do.

[Arx]

Just don't fucking cheat, and dont pretend that the game is even remotely about skill if you do.

The purpose of game AI is not to be good at the game (although it can do that). The purpose of game AI is not to push new limits in AI research (although it can do that).

The purpose of game AI is to be fun to play against.

That's the fundamental precept you should have in mind when building a game AI. If it requires it cheat, make it cheat. If it requires you lobotomise it, you lobotomise it. If it needs maphacks, you give it maphacks. There's no "purity"; there's only "fun" and "not fun".

A good friend of mine wrote a popular AI for an RTS. It cheats a ton, and can't beat many of the popular "good AIs". But it's a lot of fun to play against, and adapts itself to the skill level of its opponent. That is the only metric by which game AI must be measured. Anything else is nice, but ultimately secondary.

It's like if you buy a toaster that can load itself from a loaf, auto-butter, and send you a text when its done... but it barely actually toasts the bread. Or you buy a toaster that takes practically no space and works super fast, but also always chars your bread a bit. These are not good toasters.

Sure.  But that is a different discussion.

The initial topic is about why the hell one would call that function/process an AI when it is clearly not ever intended to be one.

It's machine learning, which is sort of an intelligent behaviour. The other thing is that neural networks are, at least in origin, based on mimicking the human brain - and what else do you call an artificial brain? Also people like calling it AI. I don't know anyone in the field who actually calls it anything other than machine learning, though.