For the record, the original Thief and Thief 2 must have been programmed by time travelers because the guard AI was designed to have different alert levels based on the number of times the guard had noticed the player. If you upset them too much or too often, they would remain in a more alert state which made them more aware of their surrounds and more aggressive in pursuit of errant noises.
Haven't played the most recent Thief game to compare, but it sure doesn't sound like much of an advance from the originals.
The article is true for AI in most games nowadays, but if we're looking at AI in a game like chess or Go then the situation changes. In games like those with both no random elements and no hidden information, duplicating the current state of the game and running many simulations of how the game will play out given a certain move becomes a very effective and very computationally intensive tool; they will benefit strongly from raw horsepower because it means that they can simulate more games in a given amount of time.
When you throw in random elements, hidden information, and/or nerfing the AI intentionally so that it behaves like a human, the benefits of these traditional machine-learning-based approaches dwindle. It's still viable for a lot of digital adaptations of modern board games, though.
In regards to AI development, it's incredibly easy for a developer to overestimate how effective AI will be before he codes it. I'm recalling the story of Trespasser here: the developers had an intricate AI system which modeled each dinosaur's emotional state that they couldn't get working in time for launch, so they ended up just cranking all of their aggression to the max and leaving it there. It's a hell of a lot easier to imagine those systems working than it is to actually get them working; I'd expect that some of that blindness is responsible for a lot of the grandiose statements that good AI requires more horsepower: the systems they're planning on implementing are much more complex than the ones that actually end up getting implemented.
And unlike graphics, AI (outside of the machine-learning approach I mentioned earlier) doesn't scale easily with computational power. You can double the size of textures or the amount of polygons in a model or do better lighting calculations relatively trivially, but adding in states and sensory data for AI adds much more in design complexity than functionality (especially since AI is entwined with gameplay development and graphics aren't).
It is really sad that today's (combat) AI seems so frigging stupid. Sure, as you pointed out making them strong is easy but strong and smart are also completely different. Why does the majority of today's AI feel that bad?
Were I to guess it is because the design goals of single player shooters have changed. If you want to make a new Call of Duty then you have a lot of design goals built around the idea of set piece battles with predictable sequences of action all striving for making a video game that plays like the highlight reel of a half dozen different action movies. The problem you have in this case is that AI that can actually make decisions beyond the most basic of shoot or not shoot lends unpredictability to the whole equation. While one conceivably write an AI that meets the exacting needs of a particular shooter sequence, it would be difficult (to understate things tremendously) to pull off and programmers don't exactly work cheaply. The solution, then, is to have most of the mooks you run into have only the most basic AI - they are generated somewhere the player can't see, they follow a route between a set of nodes, and then they take position and take occasional shots at the player until they're inevitably dispatched. For more complex behavior, you simply hard code a certain behavior - you don't want that exciting knife fight sequence to be missed because your AI decided jumping off a railing onto the player was a dumb idea given it's situation. The result is simply that most modern shooters rely very heavily on scripted sequences of actions rather than on any kind of system used to make a reasonable or rational decision.
But all is not lost - other genres find that having challenging AI that can surprise a player is important. RPGs like Skyrim are too "big" to use the tedious handcrafting that scripted sequences require and thus they rely on incredibly complex AI to run the show most of the time. Of course with Skyrim you can easily see one of the problems with leaving just about everything up to the AI - many of the common bugs in that game are the result of AI not acting like it should. Strategy games are generally simply too complex to have pre-programmed moves plotted out in advance and thus they must rely on AI to actually make decisions. As Shamus said much of the genre lives and dies on the quality of it's AI. Indeed, I think you'll find AI alive and well in most genres as the shooter is one of the few that is simple enough and small enough to rely heavily on hand-crafted sequences. And even then it is only in the big blockbuster games - ARMA and other simulators rely extensively on AI across the game.
Unless the act of motion itself is a key challenge for the player (say a Racing game or Mirror's Edge) I see no reason why the AI ought to have any trouble with movement given that any player of reasonable experience won't have to fumble around with their controller or keyboard to move left or right. That said, I do recall an Unreal Tournament mod that included lots of little realism tweaks to movement that would make your suggestion reasonable. After all, a player could trip when strafing (but it appeared to be chance based) or going down stairs (it was more common if you went backwards or were strafing) so why not shackle the AI with the same. Of course that game also enforced the idea that a gun occupies volume and many an attempt to turn and shoot at someone behind me was thwarted by a rifle smacking into a wall.
What I think it really boils down to is that the AI ought to have the same limitations as a human player. Making them bad at movement in a game where it can reasonably be expected that players will have no trouble does not make the AI seem more human.
A good topic to discuss. Sure, the combat AI routines are peanuts compared to the requirements of graphics processing but when I think of "Super AI", I'm thinking of the limitless potential that the field has yet untapped.
Think of an open world game where every single NPC has dynamic interactions with others AND the player, where their actions are determined on the fly depending on ever changing parameters. You'll have to write a human simulation, much in the vein of The Sims. When you increase the complexity and reactivity, the computational needs would surely skyrocket no?
An example:
Bob is a farmer somewhere in Tamriel. He has 2 daughters in their teens and a troublesome son who's running with the wrong crowd. He worries about whom he should marry his daughters to and how he should straighten out his son. After all, last month he had to bail him out from the local guard after his son tried to steal a noble's purse as some sort of rite of passage at the thieves guild. Thing is, the bail cost him his savings and then some. He's now in debt and all he can hope is a really good harvesting season. One of his daughters is in love with a local banker's assistant, but Bob isn't sure if the boy has the smarts to make a career and provide for his daughter should they marry. With so many worries, and the death of his wife some 6 years ago, he's been starting to have a drink before bed just so he can fall asleep better.
This is just One character among thousands in this imaginary game. You'd need to simulate a working economy, personality traits and relationships between characters and more. I wonder how impossible this kind of thing would be in terms of processing power, let alone to design algorithms for.
No.. the siege tanks are on auto fire. The programmer of this script knows how the siege tanks AI pick targets, and predicts how to avoid splash.
If you move one zergling to a tank, the tank will shoot it. If you move that one zergling with another 99 behind it, it will still be shot at as it's the first in range of the siege tank.
If a human was micromanaging those siege tanks the script wouldn't work, as it cannot predict which zerglings will be shot based on the siege tank's AI.
Just one more game the likes of Unreal Tournament 2004, Timesplitters: Future Perfect, etc. Just so I know it's the last. So I can tell myself that this is it, and so I don't keep waiting...
I wonder how the AI in Shadow of Mordor (Assassin's Creed: Middleearth) will actually work. They are making all kinds of promises in their promotional material but we'll see how much of that is real and how much of it ends up being bad.
In any case, very good story. I still feel skeptical when someone mentions X will make games better, in one way shape or form. My X being the cloud, or cloud servers. I still don't fully comprehend the matter, and an explanation from anyone is welcomed.
This is why I like games like AI War: Fleet Command and Spellforce. Both make a deliberate effort not to have the AI try to pretend to be human at all. In AI War, the AI players follow strict rules that are known to the player. The trick is that the situation is massively asymmetric and the AI could easily wipe the floor with you if it just threw everything at you, so you have to carefully plan your strategy around the AI's known reactions. Similarly, Spellforce is fairly similar to Warcraft in principle, with the human player collecting resources and building up a base and army and the AI superficially looking similar. But the AI doesn't actually build an army, it simply spawns units on a timer, and continuously escalates throughout the game. For the most part it plays fairly similarly to other RTS games, but it means you can't just turtle up in your base and fight a war of attrition, because you are guaranteed to lose if you don't make the effort to take out the enemy bases.
Dragon Age is probably also worth a mention for doing it the other way around - the human player can essentially program the AI for their party. There are only limited tools to do so, but it can give you a situation where instead of a human fighting an AI pretending to be another human, you have an AI fighting a human pretending to be another AI.
Long story short, AI that mimics humans can be tricky, but isn't always necessary at all. You don't need all players to play in the same way if the situation isn't symmetric in the first place. And since very few single player games actually are symmetric, there is plenty of opportunity for this sort of non-human AI to be used.
Bad Jim said:
Ever played a 4X game like Civilization and found out, after spying, that the AI factions are all at about the same tech level, even the tiny ones that are barely surviving? Doesn't it annoy you that the game is obviously giving them tech at certain intervals rather than make them do proper research?
Somewhat ironic that your post was about people not actually realising what the AI does, since Civilisation (the latest one at least, I can't remember exactly how previous versions worked) is one of the ones that actually doesn't cheat as you suggest. Requirements for tech (plus culture and other things) are scaled to the number of cities, precisely so that all civs generally stay at around the same tech level and simply expanding as fast as possible doesn't guarantee victory every time. The bonuses and penalties the player and AI get depending on difficulty levels are explicitly stated (mainly just happiness and base research requirements), with no other cheating occurring.
It is really sad that today's (combat) AI seems so frigging stupid. Sure, as you pointed out making them strong is easy but strong and smart are also completely different. Why does the majority of today's AI feel that bad? Did the publishers at one point just cut the AI budget in half or simply froze it while all the other costs exploded?
I think the biggest problem is that the better the AI is, the more difficult the game becomes! As a player, you do want to be able to defeat the opponents and part of that is figuring out their behaviour. As they get more complex, outwitting them becomes more like outwitting an actual person.
Also I wonder if there isn't an uncanny valley effect with AI as there is with animation - an almost intelligent opponent is a lot freakier than one who just walks backwards and forwards between two fixed locations.
Excellent article. I had not considered that it was the complexity of programming holding back AI and not processing. In that event, my only hope for the 8th gen would be that we'll see a substantial improvement in physics engines. Object texturing and such should also get a boost thanks to so much more RAM being available.
But how would you respond to claims from development studios that claim the AI is intensive?
Don't get me wrong, Titanfall is published by EA who has been trying all sorts of shenanigans to shoehorn cloud gaming so I would expect coverups but is this a blatant lie?
"For starters, it wouldn't allow for the resource-intensive AI-controlled combatants and busy battlefields the team had in mind."
Now, what I could read from that is that they needed the processing to take some of the load off of the consoles, but they say it is intensive.
dolgion said:
This is just One character among thousands in this imaginary game. You'd need to simulate a working economy, personality traits and relationships between characters and more. I wonder how impossible this kind of thing would be in terms of processing power, let alone to design algorithms for.
That's kind of what the article is talking about. Basically, it's a list of logic statements with minor calculations. Each one does require processing in much the same way that a word document takes up RAM when open, but the amount is minor.
In any event, just like you can keep opening word documents until your pc runs out of resources and crashes (or you run out of available handles), you can absolutely run into a hardware limiting number of things to keep track of in addition to what kind of game it actually is. But all of those things aren't being rendered in real time. It's all basically just a series of lines in a text document getting processed behind the scenes. This is still not difficult for machines to do. Rendering would be.
What the next gen of technology should offer is a better platform for more accurate and robust physics engines. That will make a larger difference in immersion than more polygons which have gotten pretty damn good already in the 7th generation. I do wonder about AI and what can be done to make it more streamlined like physics engines allow. For example, the Source engine can be used by a variety of games for its physics (albeit ancient), but AI can be so game specific that I wonder if it will ever be possible to have standard AI models for developers to use that can be improved on as time goes on.
In any case, very good story. I still feel skeptical when someone mentions X will make games better, in one way shape or form. My X being the cloud, or cloud servers. I still don't fully comprehend the matter, and an explanation from anyone is welcomed.
Sure, and thanks. Mainly, there's been a lot of talk about how having access to the cloud could enhance gameplay. I remember that there was mention of this with the upcoming game Destiny by Bungie. My question is how do they think that access to the cloud could accomplish this?
There's another way to categorise AI. AI that uses a few if/then statements to decide what to do, and AI that actually tries to calculate the best move.
The first kind is what nearly all current games use. It's cpu efficient but it is limited to the tricks that the programmer will teach it, and there are always exploitable holes in the logic.
The second kind is how chess AI works. It still makes mistakes, but it is much less exploitable and can potentially be much smarter than the programmer who designed it. The catch, of course, is that its' effectiveness depends on the cpu cycles you can feed it.
Galactic Civilizations 2 does actually attempt to analyse the game like this, and it is noted for its' smart AI. And yes, that AI still has many flaws, but it's also considered to be miles ahead of other 4x games.
I was going to bring up Gal Civ 2's AI, but once again, I'm beaten to the punch.
Instead, I'll expand on the design philosophy presented: different AIs do different things, but all AI behavior can be described in two ways: Proactive (Strategy) and Reactive AI (Combat AI). Which makes sense, since that's how gameplay is designed (barring games of chance).
Aimbots are purely reactive. They don't really pre-aim or funnel, or really even flank because they don't have to.
They're going to kill their target the moment the game mechanics permit it. (The Loque bot from UT99 was infamous for this; if it gets a sniper rifle, it will win every encounter the second it can see you, without fail)
Another reactive AI is the "auto-dodge" AI.
I see these a lot in fighting games, but they also occasionally pop up in shooters and vehicle games.
(like strategy AI, including Chess AI) makes "rational" decisions based on a win-condition algorithm.
Its strength is directly proportional to how many steps it can calculate ahead of its opponent; just like a chess master.
But for strategy games, there's an even better option: more "human-like" AIs (which are rare).
These are AIs that are able to think more laterally, and adapt to "chaos" by establishing both short term and long term goals within their means.
Chess AIs are completely algorithmic and very reliable because a regular Chess match always starts from the same known initial condition with all variables known. (both players can see the entire game state at all times).
But procedurally generated maps (like those found in most strategy games) generate a large number of unknowns.
Most games get around the issue of unknowns by letting the AI see everything at all times (cheating).
But to date, I've only seen two Strategy game AIs properly achieve this with minimal amounts of such cheating: Galactic Civilization 2, and Master of Orion 1 (yes, 1. Not 2 or heaven help me, 3)
Gal Civ 2 kind of cheats its first pass of choice-elimination; basing some moves on information that a regular player would not have (location of quality stars/enemy bases) but for the most part bases its actions only on known information.
Best I can tell, this is to give it a leg up in the early stages of the game, where critical developments occur and as empires grow and expose each other, this becomes less of an issue as time goes on.
Bumping up the AI-Processing option basically bumps up how many steps ahead it calculates. Still pretty algorithmic, but the result is more rational and human-like since a number of those moves are based on short term goals.
Master of Orion 1 is pretty damn savvy.
I've had games where the AI not only correctly identifies strategically valuable targets (in and out of both its empire and mine), but it pushes for them in every way that it can. If it cannot attack, it tries to sabotage or bargain. If it can do neither, it accumulates the force necessary to take it.
And when I mean targets of strategic value, I don't just mean weak targets or pricy targets with high production/population, but also targets that would give it the reach it needs to attack more vulnerable planets, OR secure its established territorial goals.
If you take a planet from the AI, the AI may not try to immediately retake it, or the AI may try to spend everything it has to retake the planet;. It all depends on whether that planet is important to its short term goals. Sometimes, the AI will only fend you off in cold war mode so that it can quietly stomp out a weaker empire elsewhere, or sue for peace with you to draw you into a conflict against an otherwise insurmountable foe.
I'm stunned at how well the AI performs in a game that old.
For comparison, I can win every game of Civilization 3 with impunity despite the MASSIVE list of cheats it can give to its AI: Omniscience, crazy Resource/Production/Research boosts and starting with multiple builders to my one. (early game production multipliers, effectively)
Hell, I've won protracted conflicts from distant 3rd place just by establishing two far off colonies, and shuffling soldiers in and out of garrison near them, the enemy armies would go on a global wild goose chase changing the target of their attack because to it, the ungarrisoned city was weaker and thus the highest priority of attack.
It didn't matter that such remote colonies were of little value compared to my core; the AI is easily destroyed by its inability to establish realistic short-term goals.
And all that ignores all other non-AI exploits like production-placeholders. With those, there isn't really even a game past the opening plays.
VoidOfOne said:
ForumSafari said:
VoidOfOne said:
In any case, very good story. I still feel skeptical when someone mentions X will make games better, in one way shape or form. My X being the cloud, or cloud servers. I still don't fully comprehend the matter, and an explanation from anyone is welcomed.
Sure, and thanks. Mainly, there's been a lot of talk about how having access to the cloud could enhance gameplay. I remember that there was mention of this with the upcoming game Destiny by Bungie. My question is how do they think that access to the cloud could accomplish this?
The claim of cloud processing, as an idea, is simple in theory:
Access to remote processing could assist games by offloading some of the processing load from the local system.
The problem with this claim, is that video game processing is overwhelmingly TIMELY.
That is, a lot of numbers need to be crunched in a relatively short period of time.
Games have other assets which require far more processing cycles than others, and come with similar but different processing arrival schedules. But compared to graphics/rendering, these assets are trivial; hence why most optimization efforts are put into the visual components.
On top of that, Cloud Processing requires a network connection (Internet, in this case), and that means the strong potential for lag. Since games are so timely, even a TINY amount of lag will have a very noticeable impact on the process.
There are a few tricks to try and mask this issue and keep the game optimized, but the skinny of it is that unless the end user has an incredible connection to the Cloud (exceptionally low latency; we're talking under 20ms at least) the Cloud isn't going to help much if at all.
AI processing, as described in the article, isn't really that resource intensive. Sure, it's a complicated decision making process relative to us, as people, but computers only cares about crunching numbers, not concept.
The math requirement of rendering involves a lot of "big" numbers; far more than any AI decisions, and exponentially more for even a minor increase in graphical fidelity.
Incidentally, AI programming, though trivial, is also timely since the computer's decisions can impact the game state in real-time as well. Meaning it too would be a poor choice for Cloud Computing, though not nearly as bad as rendering.
So not only are there few functions that Cloud Computing would significantly benefit, but the biggest one that it could help is too timely to reliably work over an internet connection.
Another way to look at it: Cloud computing is most beneficial for solving big problems slowly; like scientific equations with large degrees of accuracy (huge Taylor Series, if you've dabbled in Calculus). Equations that are easily broken into smaller chunks and are given relatively long amounts of time to crunch.
Which is the diametric opposite design virtually all video games employ, by necessity.
I remember that there was mention of this with the upcoming game Destiny by Bungie. My question is how do they think that access to the cloud could accomplish this?
'The Cloud' is basically highly available computing as a service. I'm not sure how Destiny was going to use it but it's sued by other games to outsource computation to more powerful computers and to handle saved games. The saved games storage is easy enough, that means you can pick it up from anywhere, but the more interesting idea is the computation in the cloud.
Basically a console isn't a terribly powerful platform all told and modern games have a lot of calculations in them. By moving physics calculations, AI behaviour or other stuff like that out to a more powerful series of computers they can carry out the computation a lot faster than you could on your xbox and just push the results back to the console, allowing the console to render more complex visuals or do other stuff with its' resources.
The other thing that the Cloud could be used for is less a benefit of the Cloud and more a benefit of client/server architecture generally. When you run a multiplayer game generally the content is being hosted on someone's console...who probably has the same slightly shitty Internet connection and hardware as everyone else, but who is now playing their game AND hosting the map for everyone else. By locating the server for the maps in the Cloud, automatically moving the server instance to the datacentre with the best average connectivity to the players and moving a lot of the calculations off of the consoles and in to the server they can ensure a better average connection since they've got enterprise-grade connections and a faster game seeing as they're calculating things like bullet drop and stuff for you. You're therefore computing less on local machines and sending less back up your home connection, meaning your connection is probably more stable and your console can use its' resources for better things.
Not to rain on your parade, but it really won't be.
Apart from centralizing some multiplayer processes (which is a staple of dedicated/official servers already), it won't really improve much of anything.
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
