[Cushman]

Visual quality is an "easy" problem. We know exactly how it ought to work (ray tracing has been photorealistic for some time now), it's just a matter of tweaking it to run fast enough on the hardware we have. And we're really close to pulling it off.

AI is, by contrast, a hard problem. We don't have a very good idea at all of how to exactly simulate a personality, but it's likely to be orders of magnitude more complex than the most elaborate physical simulation ever designed. The most intelligent NPC ever developed was really, really stupid. The appearance of realistic characters in games at all is an elaborate stage show. You can spend artist time making the show seem convincing in more circumstances, but evolving it past that at all is one of the hardest problems in computer science. We're nowhere near realism.

Now, with that said, in the demo they showed off some of their pathfinding and destructible environment improvements. There are plenty of gamers who care just as much (or more) about that stuff.

[gavanwoolery]

As someone who has worked intensively both in the fields of rendering and AI, I can say there are easy and hard problems in both. There is much in the way of AI that we already know, it's just that your average developer actually knows very little about AI (not trying to be insulting, it's just been true in my experience) -- and much more so your average company or publisher does not care to fund AI beyond enemies that can throw themselves into your gun barrel. Propositional logic, expert systems, backwards chaining, and neural networks are very powerful tools that we know a lot about, but most games rarely implement (perhaps trivially at most). We still tend to hand-design/script quests, when these fields of AI could provide very powerful emergent gameplay. What it mostly boils down to (for the sake of interesting gameplay) is proper knowledge-representation (i.e. abstraction) and reasoning (a prolog-like reasoning system would be a good start). And of course, path-finding/locomotion is pretty much a given. Interestingly, it is less about implementing these and more about designing a game that can use them properly - almost without doubt you need to create a game world that functions on its own, where every NPC has its own set of motivations, which is a very different task from creating a game that is limited to what the player does.

Other than that, there are still plenty of hard problems left in rendering, especially with regards to effective procedural generation (which arguably falls more under modeling). Our ray tracing is still far from "realistic" - take a more complex scene than a table with a wine glass, and we can still usually discern the difference, especially with complex materials and subsurface scattering (human skin STILL does not look quite right). As it is, we are only able to calculate light travel by points (we use the term ray tracing, but typically every ray does a number of steps, unless every model is mathematically calculated). Light travel should be determined by complex volumes -- which is pretty much impossible to simulate on today's hardware ( O(n^6) magnitude at least - growing by volume, travel distance, number of objects, volume bounces (dear god), and number of pixels calculated ).

[Cushman]

I hope I didn't seem to be implying there weren't hard problems in both fields-- but I do think there's a bigger gap in the current state of the art than you claim.

The path to realism in graphics is pretty clear because the real-world behavior of light is well-understood. There may be equations yet to be perfected, and challenges in efficiently simulating that behavior, but the direction is never in question; you can always tell if you've made it look more realistic or less.

By contrast, AI in games has reached a local maxima; the combination of pathfinding, scripted cues, an elaborate finite state machine, and a few basic heuristics is capable of simulating human behavior well enough, and with few enough glitches, that it presents a realistic universe to the player so long as they don't exceed the scripted bounds. However, if they do exceed those bounds, the whole thing appears paper-thin. You can extend the bounds with additional effort in scripting, but not infinitely, and with diminishing returns-- and doing so gets you no closer to having real AI which would be able to make those same decisions on its own.

Striking out and attempting to make an intelligent NPC-directed world from scratch... I'm not saying it's impossible, but I haven't seen anything, from AAA titles to indie games to tech demos, to imply it's coming soon.

But maybe that's just because I haven't seen the Minecraft of AI-driven gameplay yet.

[gavanwoolery]

Good points :)

I am actually working on such an AI system, believe it or not (As I am sure quite a few others are)! I don't doubt my algorithms, because I have already prototyped them and they work (at least well enough to create some interesting, if flawed gameplay), but I do doubt that I will ever get a game out that uses them effectively (making even a simple game is incredibly hard, and I have a bad habit of creating overly-complex stuff).

Your traditional AAA game does, as you mention, basically use finite state machines to determine enemy behavior (and incredibly simple ones, at that).

Why is a FSM bad?

Because someone MUST explicitly declare every combination of situation, and the paths between them.

How does my system work? It uses a combination of AI techniques that have been around a long time (just never used effectively in games).

1) Rules and queries (like prolog), with backwards chaining and propositional logic (basically a big, graph-based system). How might this work? Let's pose a few fake rules and a query. A) apples are red B) oranges are orange. C) apples and oranges are fruit D) red fruit is good...QUERY: are apples good? Using backwards chaining, this can easily be deduced.

2) Motivators and deterrents. Every NPC has their own set of motivations, as well as "bad" things they try to avoid. They attempt to maximize their output towards attaining their goals, prioritized with various points (i.e. get food: 1 point, stay alive: 10,000 points, etc)....the best outcome can be found with any traditional AI system (like a minmax tree, A*, or whatever is appropriate for the given case). This is not just about food or living, but can be emotional concepts like "seek happiness" or concepts that involve the physical world "find shelter"...

3) Basic AI for sensory and locomotion purposes (line of site, perception, planning, gathering knowledge, etc.)

4) Abstraction of communication -- not at the syntactical level, and not at the grammar level, but in its crudest terms ("higher level language")...I call it caveman-speak but there has to be a better phrase...i.e. "Me Hungry" or "Why You Steal" -- using sentence fragments, you can construct a variety of phrases without giving the computer a hard time interpreting it. On the surface, you can throw in the syntactical sugar to make it sound less crude to the user, but on the backend writing a full-blown english parser is not a wise idea.

5) There is more, but I won't type it out here...

In the simplest terms, think of it like an empty stage in a play. There is no director (Finite State Machine) but there are characters and props, and the characters have backstories/motivations/genetic makeup/birthrights that determine what they will do and how they will react to changes in their environment. When you have rules, motivators, and a knowledge base, you can let the AI run wild -- and producing new rules grows with linear complexity, rather than exponential (as with a FSM).

Alternatively, if you even played the Sims, imagine something like that, but with more comprehensive AI (NPCs seek to maximize their comfort levels (id) while acting within their restraints (ego/super ego)).

You can track my progress (or lack thereof) at gavanw.com...unfortunately I work a fulltime job in addition, so development is occasionally slow.

[sycren]

Very interesting, with your work on voxels will modellers soon have to model the inside of objects - eg) Rather than modelling the human 'skin', model each part - clothes, body, tissue, bone?

[gavanwoolery]

Well, traditional polygon-based systems will be around a long time, but I do plan to keep everything volumetric in my engine. However, for the most part a lot of the volumes are procedurally generated (i.e. the wood pattern inside a trunk is a function of distance and angle from the trunk's center).

[Cushman]

Sounds like the Minecraft of AI-driven gameplay... I would like to subscribe to your newsletter :)

[creamyhorror]

Sign me up as well. I look forward to small-scale deployments of more 'genuine' AI in games.

[gavanwoolery]

That makes 2 subscribers ;)

[scorpioxy]

Make that 3.

I've always wondered why game rules have to be so rigid (use a state machine) and always thought it was the technology but what you described makes sense and makes use of older technology but in new applications.

I just took one AI course while in college but got hooked ever since.

Question: Can you expand on the motivators and deterrents concept? Specifically, how do you relate something like "getting food" with the world objects? Assigning points to specific world objects(apple, cow, chicken...) seems to get us back to a state-machine like game. Humans have experience (memory and patterns) to guide them, what would the characters have?

[sbierwagen]

With AI, there's also the problem of being more accurate than you want to be. A triple-A title like MW3 or Skyrim is essentially "a movie you walk through". NPCs have lines, they follow a script.

You could easily have NPCs that run a little life simulation, (like in the early versions of Oblivion) but then you have the problem of NPCs going off the script. If you accurately simulate mental states, and some villager rolls a 1 on the d2000 mental state die, then proceeds to kill all the plot-critical NPCs in their village? Now you've got a more realistic, yet completely broken, game.

[sycren]

If you were to accurately simulate mental states, there would need to be some kind of cause like the player murdering all the NPC's friends in front of them to end up on 1 rather than a randomly generated value

[visural]

It seems to me, the problem with AI as applied to gaming, is that just having "good AI" doesn't mean having the smartest AI that can kick the player's butt, it means having AI that the player can observe and formulate strategies against. If it doesn't behave in a human way, and make it's decisions clear to the player then it's not fun.

In short, for gaming the measure of good AI is "fun", not "intelligence".

[wisty]

Also, better rendering is generally better (unless you want a cartoony cell renderer).

Better AI might not always be better. You don't want stupid AI bugs, but you don't necessarily want the AI to be smarter than the player. You may want an AI with a realistic character, but maybe not. Good AI makes some sense in a sandbox world, but it's not necessarily fun.

[lloeki]

> your average company or publisher does not care to fund AI beyond enemies that can throw themselves into your gun barrel

Ironically the game where I've seen the most convincing AI currently is the Forza Motorsport series (increasing with each release). There is nothing short of impressive in seeing opponent drivers trying to pressure you to fault, intentionally breaking on you way after the apex just when you are about to floor it, showing excess of confidence throwing them out of the track, or reacting to pressure you put on them (you can literally see some panicking or getting aggressive in their driving), learning to clock faster times lap after lap and race after race, and even from your own lines. Sure they're nowhere near a real human, and with experience you can land better times than the AI and outsmart them, but there's plenty of physics and graphics going on already and there are trade-offs to be made.

[gideons]

In a video game anything the player doesn't see is a waste of time. In the case of games with large branching story lines it's a tradeoff between putting effort in to content which might be unseen and adding "depth." AI is the same. A genius AI that does all kinds of magic in the background is useless if the player never gets to see the magic.

Games don't want true AI solutions, they want to entertain the player. Put enough veneer on the world to make the player suspend disbelief and move on. Anything else is a waste of money for the most part.

[InclinedPlane]

Static photorealistic scenes are still a pretty hard problem. If it was just "ray tracing and you're done" it'd be a different story. You have complex lighting interactions (best modelled using the radiosity technique) and you have complex sub-surface light interactions which are a prominent compnent to making ordinary human skin appear real. These aren't even completely solved problems in the state of the art today.

Now, all of that is just square zero. A mere preface to the real problems: realistic animation and realistic interactivity. You can make a pretty decent photorealistic world with current technology but it would have to be an untouchable, unmoving, unrealistic frozen world. Adding in animations is a problem at the same computational intensity as the rendering problem. And making the world capable of being interacted with a level of fidelity approaching the quality of the rendering and animation is a problem so lacking in proposed solutions that it doesn't even have a sound theoretical underpinning yet.

[dualogy]

Classical raytracing by itself is highly unrealistic unless you also implement an add-on physical-based rendering-equation-solver such as Path Tracing or Photon Mapping. The first one beginning highly noisy and "converging" to photo-realistic over a looong time -- the second being only as realistic as the number of photons your hardware can manage.

Now without such a physical-based rendering-equation-solver, realtime raytracing of "simple scenes" (couple of spheres and boxes) is doable today. The more complex your scenes get (even with acceleration structures), the more you need to lower your viewport resolution. Your high-end desktop GPU can then handle raytracing for a low-end "mobile" resolution such as 320x240 or whatever. But your low-end mobile GPU probably can't. Also for fully accurate antialiasing you'd need to render at 2x your resolution and downscale.

tl;dr: realtime raytracing of simple scenes is feasible, realtime photorealistic full-screen antialiased raytracing of complex scenes isn't and if current tech development trends continue, won't be for another decade perhaps.

I don't see "we're really close to pulling it off" because as GPU flops and caps happily increase, so do unfortunately screen resolutions. In realtime raytracing, we're always lagging behind and always have to compromise: (1) high resolution, (2) antialiasing, (3) complex scenes, (4) photo-realism -- for real-time speeds, pick any two.

[aidenn0]

I didn't read the grandparent as saying we are really close to real-time path tracing; I read it more as real-time rasterization is getting rapidly closer to looking like path tracing both in the quality of the scenes and the types of scenes it can represent.

Also FLOPS has been increasing much faster than resolutions. 1920x1200 is about 3x the pixels of 1024x768. Graphics FLOPS has increased way more than 3x since 1024x768 was considered a high-resolution screen.

[dualogy]

Good points ... that's a silver lining, too =)

[dclowd9901]

I think the fact that strides have been made in the general area of dynamicizing certain actions (like simultaneously rendering mouth movements with spoken language in games) are steps toward that grand goal of games having AI that not only acts smart, but even has dynamic scripting, based on "observances" and other environmental stimuli. Also, hinted/dynamic animation is pretty unbelievable now. Just see how real it looks when a character is tackled in a Madden game.

I'm confident even the "hard" problems will be solved in the very near future.

[Cushman]

That's a good point, and maybe I was too binary in my original statement-- there are a lot of parts of NPC design that are more physics simulation than AI. I'd still say there's a categorical distinction between "respond appropriately to being hit by a linebacker" and "respond appropriately to someone putting a bucket on your head", though. Like you say, we're close to realism on the former; much farther from the latter.

[arketyp]

>Visual quality is an "easy" problem. We know exactly how it ought to work (ray tracing has been photorealistic for some time now), it's just a matter of tweaking it to run fast enough on the hardware we have. And we're really close to pulling it off.

That's a good point. And I don't even think it's so much about AI being really hard as it is about photorealism being such a clear goal. Because really there is no limit to the other dynamics that can be explored that aren't as unattainable as a human AI. I still think the ghosts in Pac-man are clever. And while I'm by no means tuned in on the gaming community today, just the fact that you still play these super good-looking games with a few buttons and two joysticks says a lot.