Is an ingame A.I. revolution coming?

[moonbogg]

They had an A.I. on a USB drive that beat the world's best Dota 2 players, so I'm wondering when video games will incorporate real A.I. that behave like people would. It could change so much and add amazing replayability. Imagine not being able to tell the difference between an A.I. and a real person in the single player part of a game? Please bring it.

 

[XavierMace]

That's not happening any time soon.

 

[moonbogg]

Well crap. Fine then.

/thread?

 

[DaveSimmons]

Nope.

"Deep learning" neural nets created by GPUs and cloud clusters have gotten much better at pattern matching, which maps to things like language translation and image recognition. But the AI does not really "understand" anything. It doesn't actually think. It can only "plan" in very limited and tightly constrained cases where it is doing a simple search of moves.

 

[werepossum]

Nope.

"Deep learning" neural nets created by GPUs and cloud clusters have gotten much better at pattern matching, which maps to things like language translation and image recognition. But the AI does not really "understand" anything. It doesn't actually think. It can only "plan" in very limited and tightly constrained cases where it is doing a simple search of moves.

But for gaming we don't really need AI that can think. We need AI that can simulate thinking.

 

[DaveSimmons]

But for gaming we don't really need AI that can think. We need AI that can simulate thinking.

True, but it requires an awful lot of simulation to reach this level from the OP:

Imagine not being able to tell the difference between an A.I. and a real person in the single player part of a game?

That's passing the Turing test except with a 3D avatar speaking and taking realistic actions instead of just creating text answers to text questions.

The NPCs in existing single-player games use a combination of scripting, pre-recorded voice acting and motion capture to sort of do this, but that behavior is on rails.

 

[Zenoth]

Whoever coded the A.I. back in the original F.E.A.R. must have definitely had good scores in 'coding school'. Some games already had pretty good A.I. years ago, they are rare gems but they did happen. I also remember the Marines in the first Half-Life, compared to all the rest of the enemies' A.I. in the game for some reason the coders decided to go coding bunkers for the Marines. Now, of course, some of those old game examples are... well, old today. You'd look at the Marines in H-L and immediately see their rather slow reaction time and movement in response to threat, etc. Well back in 1998 hardware wasn't the same either.

A decent modern example is in none other than Breath of the Wild, some of the wildlife has pretty good A.I. and react to a ton of conditions from the player and the environment. It's a very small detail but I can throw a rock at a butterfly and contrarily to how it'd "behave" in most other games (I.E. stay there or keep playing its static animation in place regardless of what I do) it flies away and, if lucky enough, it lands some distance away and goes back to its normal idling 'routine'. That, to me, is a sign of care, passion and dedication on the coders' part. It's not that it's anything "complex", per se, it's just that they code a LOT of different types of reactions to different types of conditions in a (imaginable) large list of behaviors. And, ultimately, it is the PLAYER who does the "simulation" part with suspension of disbelief, it goes a very long way in 'complementing' (not quite replacing) actual A.I.; but if you can immerse yourself in a game (like I myself can in Breath of the Wild) you soon forget how 'better' it could be if they worked on A.I. alone for another 10 years and just play a good game.

I don't know the technical shenanigans behind it much but what I do know as a gamer who started on the Master System and grew up with the NES, SNES and Genesis is that to this day there's been only a very small number of games that demonstrated a true, undeniable milestone 'leap' in A.I. behavior (be it in the environment from enemies or A.I. bots or what have you). To come back to F.E.A.R. for an example of such cases, just pointing out that it truly had an impressive A.I. from the enemies (well the enemy soldiers, a bit like the Marines specifically in the first Half-Life). Sometimes good A.I. also come in the form of heavily scripted encounters, it really doesn't HAVE to be 'dynamic' or reacting in real time "like a human would". I can think of scripted Boss encounters in some games, where the bosses would have a set of moves but the A.I. would go with the most appropriate set against a specific attack you'd do, either alone or in a group (happened in some MMORPGs, not sure about offline games but I suppose that at least some specific bosses in series like Dark Souls had something along the line going on at some point).

The gist is that at least for me, "good A.I." is very broad and contextual, and also depends on the actual 'need' of the game for or against the player (I.E. does the game NEED to have good A.I. in it? If it's friendly, relaxing pure-fun game in the veins of Mario Bros then... no, I don't want Goombas making Alexander The Great level plans to destroy the Mushroom Kingdom).

 

[BSim500]

They had an A.I. on a USB drive that beat the world's best Dota 2 players, so I'm wondering when video games will incorporate real A.I. that behave like people would. It could change so much and add amazing replayability. Imagine not being able to tell the difference between an A.I. and a real person in the single player part of a game? Please bring it.

As Zenoth said above, after all this time, after all these years and MOAR CORES worth of unused horsepower, ask 100 gamers who've been playing for 20 years to name a memorable AI in games and many will remember FEAR1. Hell the same developers (Monolith) put some surprising things in earlier NOLF games (enemies flipping over tables, then hiding behind them then occasionally blindly firing over the top with only the gun and the edge of one hand sticking out, reacting to light-bulbs, doors left open, tracks in the snow, different behaviour depending on whether attacking as an individual or part of a group, different behaviour depending on how injured they are, etc). According to the Wiki page, enemies react to 11 different stimuli in even the first NOLF 1 game on the primitive Lithtech 2.2 Engine (which predated Jupiter Engine (NOLF2) which predated Jupiter EX Engine (FEAR1)).

It's all down to developer attitude & intention and has been all these years. Even if the AI isn't "real", some very clever mimicry of intelligent enemies can be scripted in without it feeling like it's too scripted, and we never were waiting for more horsepower for this stuff. Sadly given the current trend towards trashy micro-transaction laden grind-fests, it feels more like a lost art with each new release, with all the talented AI coders being increasingly replaced with skin artists and DLC salesmen...

 

[Borealis7]

the Bots in the original Unreal game used to kick my ass every time. i don't need better AI opponents, i need cybernetic enhancements!

 

[book_ed]

The problem is with how much time and money developers want to invest in it and since most are doing it just for the money, progress is slow. So I wouldn't hold my breath.

 

[werepossum]

True, but it requires an awful lot of simulation to reach this level from the OP:

That's passing the Turing test except with a 3D avatar speaking and taking realistic actions instead of just creating text answers to text questions.

The NPCs in existing single-player games use a combination of scripting, pre-recorded voice acting and motion capture to sort of do this, but that behavior is on rails.

I took the OP’s point to be AI whose play was indistinguishable from a human’s, still a high bar but not nearly as high. For most games, I don’t think we will see true, unscripted AI within my lifetime - the hardware and programming requirements are simply too huge - but AI certainly needs far more assets than are allocated today. I’m something of a graphics whore, but better AI is waaay higher on my wish list than better graphics.

 

[werepossum]

Whoever coded the A.I. back in the original F.E.A.R. must have definitely had good scores in 'coding school'. Some games already had pretty good A.I. years ago, they are rare gems but they did happen. I also remember the Marines in the first Half-Life, compared to all the rest of the enemies' A.I. in the game for some reason the coders decided to go coding bunkers for the Marines. Now, of course, some of those old game examples are... well, old today. You'd look at the Marines in H-L and immediately see their rather slow reaction time and movement in response to threat, etc. Well back in 1998 hardware wasn't the same either.

A decent modern example is in none other than Breath of the Wild, some of the wildlife has pretty good A.I. and react to a ton of conditions from the player and the environment. It's a very small detail but I can throw a rock at a butterfly and contrarily to how it'd "behave" in most other games (I.E. stay there or keep playing its static animation in place regardless of what I do) it flies away and, if lucky enough, it lands some distance away and goes back to its normal idling 'routine'. That, to me, is a sign of care, passion and dedication on the coders' part. It's not that it's anything "complex", per se, it's just that they code a LOT of different types of reactions to different types of conditions in a (imaginable) large list of behaviors. And, ultimately, it is the PLAYER who does the "simulation" part with suspension of disbelief, it goes a very long way in 'complementing' (not quite replacing) actual A.I.; but if you can immerse yourself in a game (like I myself can in Breath of the Wild) you soon forget how 'better' it could be if they worked on A.I. alone for another 10 years and just play a good game.

I don't know the technical shenanigans behind it much but what I do know as a gamer who started on the Master System and grew up with the NES, SNES and Genesis is that to this day there's been only a very small number of games that demonstrated a true, undeniable milestone 'leap' in A.I. behavior (be it in the environment from enemies or A.I. bots or what have you). To come back to F.E.A.R. for an example of such cases, just pointing out that it truly had an impressive A.I. from the enemies (well the enemy soldiers, a bit like the Marines specifically in the first Half-Life). Sometimes good A.I. also come in the form of heavily scripted encounters, it really doesn't HAVE to be 'dynamic' or reacting in real time "like a human would". I can think of scripted Boss encounters in some games, where the bosses would have a set of moves but the A.I. would go with the most appropriate set against a specific attack you'd do, either alone or in a group (happened in some MMORPGs, not sure about offline games but I suppose that at least some specific bosses in series like Dark Souls had something along the line going on at some point).

The gist is that at least for me, "good A.I." is very broad and contextual, and also depends on the actual 'need' of the game for or against the player (I.E. does the game NEED to have good A.I. in it? If it's friendly, relaxing pure-fun game in the veins of Mario Bros then... no, I don't want Goombas making Alexander The Great level plans to destroy the Mushroom Kingdom).

Good points, and you are correct that I want immersive AI, not necessarily human-level AI. Let’s face it; if I am going up against a dozen soldiers with human level AI, they are going to flank and destroy me every time, period. F.E.A.R. and Half-life are two of my all time favorite games not because they had human level AI, but because they had immersive AI.

 

[werepossum]

As Zenoth said above, after all this time, after all these years and MOAR CORES worth of unused horsepower, ask 100 gamers who've been playing for 20 years to name a memorable AI in games and many will remember FEAR1. Hell the same developers (Monolith) put some surprising things in earlier NOLF games (enemies flipping over tables, then hiding behind them then occasionally blindly firing over the top with only the gun and the edge of one hand sticking out, reacting to light-bulbs, doors left open, tracks in the snow, different behaviour depending on whether attacking as an individual or part of a group, different behaviour depending on how injured they are, etc). According to the Wiki page, enemies react to 11 different stimuli in even the first NOLF 1 game on the primitive Lithtech 2.2 Engine (which predated Jupiter Engine (NOLF2) which predated Jupiter EX Engine (FEAR1)).

It's all down to developer attitude & intention and has been all these years. Even if the AI isn't "real", some very clever mimicry of intelligent enemies can be scripted in without it feeling like it's too scripted, and we never were waiting for more horsepower for this stuff. Sadly given the current trend towards trashy micro-transaction laden grind-fests, it feels more like a lost art with each new release, with all the talented AI coders being increasingly replaced with skin artists and DLC salesmen...

Spot-on. We have available hardware assets to do very good AI, just not the allocated developer assets. However, this may be due to console limitations as much as greed.

 

[Borealis7]

remember that the Game Engine loop for each frame is:
1) load all active entities
2) calculate AI behavior for each entity
3) update positions in world
4) render next frame

you really wouldn't want complex AI calculations between those 1 and 3 phases to delay your frame rendering, that would lead to degraded performance regardless of GPU power.

 

[moonbogg]

I took the OP’s point to be AI whose play was indistinguishable from a human’s, still a high bar but not nearly as high. For most games, I don’t think we will see true, unscripted AI within my lifetime - the hardware and programming requirements are simply too huge - but AI certainly needs far more assets than are allocated today. I’m something of a graphics whore, but better AI is waaay higher on my wish list than better graphics.

That's what I meant. Also, I don't know how old you are, but I expect to see real A.I. in games in the not too distant future. It may not be stand alone PC or console games at first, but simpler games with real time cloud service. The A.I. is cloud based and it interacts as an A.I. server for the active player base, learning from each player as it goes, accumulating experience from each player. So I can see future games like Battlefield using A.I. cloud services to create lifelike enemies and other features of the game for both single player and multi player.

 

[WhoBeDaPlaya]

the Bots in the original Unreal game used to kick my ass every time. i don't need better AI opponents, i need cybernetic enhancements!

UT99 - Morpheus with instagib and Godlike bots = fun times!

 

[book_ed]

https://youtu.be/p7PlQ-q17tM?t=6m48s

https://gpuopen.com/archive/gpu-demos/radeon-hd-4800-series-real-time-demos/

The Froblins demo employs state-of-the-art, massively parallel artificial intelligence computation for dynamic path finding and local avoidance on the GPU. The froblins busily move from goal to goal while avoiding treacherous regions of the terrain. The characters spend time working at gold mines, foraging wild mushrooms, and napping at their camp sites. The user can explore every corner of this virtual world by flying around the environment using a variety of input paradigms. The user may also influence the behavior of the froblins by placing new goals in the environment and even adding new obstacles such as dangerous poison fields and summoning pull-rightening ghost froblins! As new goals and obstacles are placed in the environment, the froblins adapt by dynamically changing their paths.

Many rendering scenarios, such as battle scenes or urban environments, require rendering of large numbers of autonomous characters. Crowd rendering in large environments presents a number of challenges, including visibility culling, animation, and level of detail (LOD) management. These have been traditionally CPU-based tasks, trading some extra CPU work for a larger reduction in the GPU load. However, the per-character cost can be a serious bottleneck in that scenario. Furthermore, CPU-side scene management is difficult if objects are simulated and animated on the GPU as they are in the Froblins demo. This demo uses Direct3D® 10.1 functionality to perform view-frustum culling, occlusion culling, and LOD selection entirely on the GPU, allowing thousands of GPU-simulated characters to be rendered with full shadows in arbitrary environments.

That stuff was already doable on 1TF card (HD4850) and now we have 12TF+ cards and nothing like that in sight...

 

[BSim500]

That's what I meant. Also, I don't know how old you are, but I expect to see real A.I. in games in the not too distant future. It may not be stand alone PC or console games at first, but simpler games with real time cloud service. The A.I. is cloud based and it interacts as an A.I. server for the active player base, learning from each player as it goes, accumulating experience from each player. So I can see future games like Battlefield using A.I. cloud services to create lifelike enemies and other features of the game for both single player and multi player.

It's always nice to dream but the one thing you can take away from modern games is that they're designed down to a price-point / pile it high, sell it (not so) cheap business model, and super expensive cloud A.I with high ongoing logistical costs unique to each game is literally the last thing they'd offer. A single-player game which had a finite shelf-life (ie, after a couple of years the plug on the AI servers gets pulled and the game literally becomes an unplayable digital doorstop) or a "your Internet lagged / glitched out for 0.1s, let me boot you from this single-player game" is also the last thing I'd buy.

 

[werepossum]

That's what I meant. Also, I don't know how old you are, but I expect to see real A.I. in games in the not too distant future. It may not be stand alone PC or console games at first, but simpler games with real time cloud service. The A.I. is cloud based and it interacts as an A.I. server for the active player base, learning from each player as it goes, accumulating experience from each player. So I can see future games like Battlefield using A.I. cloud services to create lifelike enemies and other features of the game for both single player and multi player.

For the first, I am 57. For the second, I think the one area where we might see an orders of magnitude jump in AI would be MMORPGs. Many people play those alone, and such AI would be a natural. However, there is another area which have to co-evolve to make a single, cloud-based AI actually simulate a human - machine language. Voice acting is fairly expensive, and at least hundreds of thousands of lines would have to be recorded to simulate a human. Luckily this is something the non-gaming development community is tackling big time. For Battlefield it would be somewhat easier since the NPCs aren’t expected to say much, but they already get human opponents who pay them for the privilege. Strategy games might be another area where one super-capable cloud-based AI makes sense. But again, if you have a good net connection, the company actually gets paid to provide real human opponents.

I think there is a real opportunity here for an indie team. With the liberal licensing plans of Unreal and Cryengine, graphics are pretty much handled. Adding top level AI programmers to the team could pay huge dividends since the major developers no longer put much capital into AI.

 

[moonbogg]

I personally think A.I. will be everywhere in the same way the internet is. It starts out slow, expensive and exclusive. Then the heavy investing starts, it spreads, gets cheaper and better and bam, its everywhere from your PC to smart phone, refrigerator, car, cat collar etc etc. I think games will be loaded with it and without A.I. it would be like playing without the internet today or trying to enjoy a game without a GPU or something. That's what I think at least. Also, it will take a long time to get there until suddenly it doesn't. I think it will be nowhere to be seen, and appear far off in the future. Then suddenly you'll hear about it and it will explode onto the scene. It will be used everywhere for everything and it will seem so sudden, you'll be asking yourself, "When did this become my world?".

/prophetic rant

 

[Denithor]

LOL every time I hear someone talking about game AI or see an Intel/nVidia/AMD commercial featuring AI features, I cannot help but think of SkyNet. That's how it all gets started, self-cognizant AI. And, if it's already in the proverbial cloud when it's established, there will be no way to turn it off if things go sideways.

 

[sweenish]

For the first, I am 57. For the second, I think the one area where we might see an orders of magnitude jump in AI would be MMORPGs. Many people play those alone, and such AI would be a natural. However, there is another area which have to co-evolve to make a single, cloud-based AI actually simulate a human - machine language. Voice acting is fairly expensive, and at least hundreds of thousands of lines would have to be recorded to simulate a human. Luckily this is something the non-gaming development community is tackling big time. For Battlefield it would be somewhat easier since the NPCs aren’t expected to say much, but they already get human opponents who pay them for the privilege. Strategy games might be another area where one super-capable cloud-based AI makes sense. But again, if you have a good net connection, the company actually gets paid to provide real human opponents.

I think there is a real opportunity here for an indie team. With the liberal licensing plans of Unreal and Cryengine, graphics are pretty much handled. Adding top level AI programmers to the team could pay huge dividends since the major developers no longer put much capital into AI.

I think you're spot on that MMO's are a prime place to start rolling these ideas out. The first Guild Wars had NPC's you could hire to go questing with you, but they were worthless. Making them too good can make actual partying less efficient, though. So the AI would have to be able to play at or around the level of the player. That alone is a completely different learning process than the actual NPC behaviors.

As far as cloud-based stuff goes, we are dipping our toes. Titanfall leverages Azure for AI calculations. Now, that AI is not very robust, it was more of a volume solution than a behavioral solution. Crackdown 3 is supposed to offload city-wide physics calculations to the cloud. So, we're getting into the idea of offloading calculations in our games. It presents its own dangers for the future preservation of these games, and I think that's something that needs to be considered sooner than later.

And as far as dynamic voicing goes, that's what our voice assistants do. Yes, we still need quite a bit of stuff recorded, but it's not the recording of specific phrases. It's the combining and smoothing of syllables with appropriate pitch and tonal adjustments. Again, another machine learning problem.

Tools are out there, but until decent ones are openly developed, we will likely not see "good AI" in our games at the scale being discussed.

 

[Charlie22911]

I don't think better AI would be hard to do at all, cellphones out today even have dedicated silicon for the task.
One way I could think of it being implemented is perhaps as part of the uncore of a CPU or as part of the SM in a GPU.