6 ms·

I wonder if AI is where games can make use of many cores. If the game logic can not be reasonibly parallelized anymore, the unused computing capacity can be use

by madez 10y ago

I wonder if AI is where games can make use of many cores. If the game logic can not be reasonibly parallelized anymore, the unused computing capacity can be used to add "intelligent" agents into the game. They don't need to do revolutionary things, just being more complex, more "reasonable" and less predictable than the usual agents would be enough. I think that has the potential to make some games more interesting. Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy. Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples?

ge96 10y ago
Mr Anderson...
dtromero 10y ago
Interested in this as well. Applying machine learning in this way could create some interesting game play.
mtrycz 10y ago
I was told (but didn't verify it myself, nor have any sources on this) that in Half Life 2 (2004), your enemies would "learn" from your attack patterns and adapt to it. At the same time, as to make the game enjoyable/chellenging and not frustratingly difficult, the agents would adjust their difficulty level to your skill. Take it as an anecdote.
- selectodude 10y ago
  That was actually Left 4 Dead. https://en.wikipedia.org/wiki/Left_4_Dead#AI_Director https://en.wikipedia.org/wiki/Left_4_Dead#AI_Director
  - umurkontaci 10y ago
    The director in L4D was had a very simple logic with inputs like number of alive special infecteds and number of zombies to control spawns. It's like seeing a health kit when your health is low.
    - selectodude 10y ago
      And the director in L4D2 actually changed layouts of levels to suit difficulty. Have to start somewhere, though.
    - Crespyl 10y ago
      IIRC it tried to be slightly more clever (might've been L4D2) in that it also kept track of the recent history of the game and attempted to balance the pace of the action. If it thought players had had it easy for a while it might throw a larger waves of monsters and specifically deny spawning health for a few minutes. There's a good presentation somewhere about how it would try to balance the rising action with periods of "stress" and "relief" over the course of a level.
- squeaky-clean 10y ago
  I've played an unhealthy amount of HL2 and spent a while modding it, and I don't believe this is true. People say things like this a lot about games with decent AI. For example, the old quake 3 hoax about bots "learning" to become peaceful on a server that was left running for 4 years. It's fun to imagine the AI as more than just a set of scripts. The Valve Developer Community has some okay documents about the AI [0], moddb has some nice articles too (no links offhand). And most of the source for combine guards is available in the Source SDK [1], you can see tons of if's and hard coded sequences of actions. I've read that the AI was changed after the Orange Box release, so maybe there was some learning features to them originally. Or maybe those features do exist within HL2's private source files, though it doesn't "feel" like it. [0] https://developer.valvesoftware.com/wiki/AI_Learning:_CombineSoldier https://developer.valvesoftware.com/wiki/AI_Learning:_Combin... [1] https://github.com/ValveSoftware/source-sdk-2013/blob/master/sp/src/game/server/hl2/npc_combine.cpp https://github.com/ValveSoftware/source-sdk-2013/blob/master...
- JimmyAustin 10y ago
  I've heard a similar anecdote about a racing game (I think it might have been in Forza) where the developer had a AI where individual racers could hold grudges and that would affect how aggressively they treated you (bump someone in lap 1 and they'll be more aggressive in lap 3). Eventually they had to take it out for various reasons, but didn't advertise that fact for obvious reasons. When the game was finally released several reviewers brought up the feature as a plus, not realising that they were just reading into the actions of the standard AI. As a side note, the writeup of F.E.A.R's AI is a fascinating read. By combining very simple goals and actions, they created AI that is still remembered as incredibly good. http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear....
gcp 10y ago
I don't know about current games (I suspect they have enough to load the CPU with gfx/physics to devote much to AI), but state of the art AI software can definitely be many core. It's very useful for anything doing Monte Carlo simulations, tree search, planning.
- johncolanduoni 10y ago
  Games haven't really been pushing CPUs very hard for the past few years. All the mucking about with overclocking has been purely a matter of culture recently, as you can play top tier games in 4K with a mid-range CPU, as long as it can handle enough PCIe lanes for your monster GPU setup.
  - zer0t3ch 10y ago
    Depends on the game, really. That may be true for some, but far from all.
    - wolrah 10y ago
      Take a look at the game sections from the recent Kaby Lake reviews. Here's Anandtech's for example: http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-91w-review-the-new-stock-performance-champion/6 http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-... Most of them look pretty much like that, with very little difference until you start getting to really low-end or old chips. Games are mostly GPU limited these days and those that are CPU limited are mostly leaning on one or two cores and thus haven't really gained much from recent generations.
      - gcp 10y ago
        You're arguing that benchmarks of recent top of the line chips (which at most tend to have a few tens of percent difference) not showing large performance difference is evidence that games aren't CPU limited, whereas the same test shows drops in single-thread performance immediately cause large performance differences (see AMD chips). If anything, this shows that games aren't strongly multi-threaded (which comes with severe development and debugging impact) or aren't made to scale beyond the highest end chips consumers are likely to have. Lastly, consider that while the graphics effects and physics effects of a game can be scaled down, scaling down the AI in many instances would cause balancing issues.
Le_SDT 10y ago
Imagine an RTS or something similar where you play with an AI and he learns your behaviours. He tries to help you. At first you deny what he does, he learns, do something else. A bit like the beast in black & white!
forgotmysn 10y ago
pretty sure Rival Theory was working on something similar to that
_John-Titor_ 10y ago
There's a recent game called Hello Neighbour where the eponymous neighbor learns from the patterns you use to sneak into his house. The one video I saw of the game looked freaky. 10/10 would play.
squeaky-clean 10y ago
> Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy. > Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples? It's a really tough trade-off between fun and realism. It's cool that your units have personality for a bit "Oh, look, those 2 dwarves are fleeing from the troll fight early!", but not fun once they disobey your intentions "That fucking troll has 3 hit points left. If my 2 dwarves didn't run away, it would be dead! Now it's destroying my base!" One game I remember striking a good middle ground with this was the original Battle For Middle Earth. It changed animations, sounds and other aesthetic things based on AI, while the game remained the same mechanically. If your units ran up against a mountain troll, you would see the ones in front of it cowering and holding up shields, while units behind attack furiously, but every unit was really doing the same damage. Or if you left troops of Orcs just standing around too long, they'd break out into fights and kill other orcs, but the animation is done as a trick where as one orc model "dies", a new one spawns in the group. So you didn't really ever lose Orcs, even though there would be orc corpses strewn wherever you had been camping. I think the main reason though is just that AI doesn't really sell. Graphics, physics and voice actors sell. You can blame a failure on AI, but very rarely is success attributed to good AI in a game.
- Karunamon 10y ago
  Bad AI often winds up panned in a game's review scores, which have an impact on sales. For instance, bad pathfinding in an RTS game can make the game almost unplayable. Halo is just about legendary at this point for teammate AI that is worse than worthless [1]. Any one of these things can destroy immersion at a critical point. AI is to video games as IT is to a modern enterprise. If it's doing its job, nobody notices, but if it's bad, it sticks out like a sore thumb. [1]: https://www.youtube.com/watch?v=vdbg8h3d3uM https://www.youtube.com/watch?v=vdbg8h3d3uM
  - organsnyder 10y ago
    Or, if it does its job really well, it gets accused of cheating. My former company had a backgammon game with a really good AI. We constantly got accused of cheating, even though the dice-roll algorithm was identical (it didn't have a parameter to specify who the roll was for). What people missed is that a good backgammon player will position themselves to benefit from a wider set of possible dice outcomes.
JabavuAdams 10y ago
Game logic has been parallelized since about 2005-2010.
BickNowstrom 10y ago
In ARMA opponent AI is a parameter. When you have a lot of players on one server, you turn down enemy AI so you don't bog down the entire server with pathfinding calculations.

by madez 10y ago

ge96 10y ago

Mr Anderson...

dtromero 10y ago

Interested in this as well. Applying machine learning in this way could create some interesting game play.

mtrycz 10y ago

I was told (but didn't verify it myself, nor have any sources on this) that in Half Life 2 (2004), your enemies would "learn" from your attack patterns and adapt to it. At the same time, as to make the game enjoyable/chellenging and not frustratingly difficult, the agents would adjust their difficulty level to your skill. Take it as an anecdote.

selectodude 10y ago
That was actually Left 4 Dead. https://en.wikipedia.org/wiki/Left_4_Dead#AI_Director https://en.wikipedia.org/wiki/Left_4_Dead#AI_Director
- umurkontaci 10y ago
  The director in L4D was had a very simple logic with inputs like number of alive special infecteds and number of zombies to control spawns. It's like seeing a health kit when your health is low.
  - selectodude 10y ago
    And the director in L4D2 actually changed layouts of levels to suit difficulty. Have to start somewhere, though.
  - Crespyl 10y ago
    IIRC it tried to be slightly more clever (might've been L4D2) in that it also kept track of the recent history of the game and attempted to balance the pace of the action. If it thought players had had it easy for a while it might throw a larger waves of monsters and specifically deny spawning health for a few minutes. There's a good presentation somewhere about how it would try to balance the rising action with periods of "stress" and "relief" over the course of a level.
squeaky-clean 10y ago
I've played an unhealthy amount of HL2 and spent a while modding it, and I don't believe this is true. People say things like this a lot about games with decent AI. For example, the old quake 3 hoax about bots "learning" to become peaceful on a server that was left running for 4 years. It's fun to imagine the AI as more than just a set of scripts. The Valve Developer Community has some okay documents about the AI [0], moddb has some nice articles too (no links offhand). And most of the source for combine guards is available in the Source SDK [1], you can see tons of if's and hard coded sequences of actions. I've read that the AI was changed after the Orange Box release, so maybe there was some learning features to them originally. Or maybe those features do exist within HL2's private source files, though it doesn't "feel" like it. [0] https://developer.valvesoftware.com/wiki/AI_Learning:_CombineSoldier https://developer.valvesoftware.com/wiki/AI_Learning:_Combin... [1] https://github.com/ValveSoftware/source-sdk-2013/blob/master/sp/src/game/server/hl2/npc_combine.cpp https://github.com/ValveSoftware/source-sdk-2013/blob/master...
JimmyAustin 10y ago
I've heard a similar anecdote about a racing game (I think it might have been in Forza) where the developer had a AI where individual racers could hold grudges and that would affect how aggressively they treated you (bump someone in lap 1 and they'll be more aggressive in lap 3). Eventually they had to take it out for various reasons, but didn't advertise that fact for obvious reasons. When the game was finally released several reviewers brought up the feature as a plus, not realising that they were just reading into the actions of the standard AI. As a side note, the writeup of F.E.A.R's AI is a fascinating read. By combining very simple goals and actions, they created AI that is still remembered as incredibly good. http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear....

gcp 10y ago

I don't know about current games (I suspect they have enough to load the CPU with gfx/physics to devote much to AI), but state of the art AI software can definitely be many core. It's very useful for anything doing Monte Carlo simulations, tree search, planning.

johncolanduoni 10y ago
Games haven't really been pushing CPUs very hard for the past few years. All the mucking about with overclocking has been purely a matter of culture recently, as you can play top tier games in 4K with a mid-range CPU, as long as it can handle enough PCIe lanes for your monster GPU setup.
- zer0t3ch 10y ago
  Depends on the game, really. That may be true for some, but far from all.
  - wolrah 10y ago
    Take a look at the game sections from the recent Kaby Lake reviews. Here's Anandtech's for example: http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-91w-review-the-new-stock-performance-champion/6 http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-... Most of them look pretty much like that, with very little difference until you start getting to really low-end or old chips. Games are mostly GPU limited these days and those that are CPU limited are mostly leaning on one or two cores and thus haven't really gained much from recent generations.
    - gcp 10y ago
      You're arguing that benchmarks of recent top of the line chips (which at most tend to have a few tens of percent difference) not showing large performance difference is evidence that games aren't CPU limited, whereas the same test shows drops in single-thread performance immediately cause large performance differences (see AMD chips). If anything, this shows that games aren't strongly multi-threaded (which comes with severe development and debugging impact) or aren't made to scale beyond the highest end chips consumers are likely to have. Lastly, consider that while the graphics effects and physics effects of a game can be scaled down, scaling down the AI in many instances would cause balancing issues.

Le_SDT 10y ago

Imagine an RTS or something similar where you play with an AI and he learns your behaviours. He tries to help you. At first you deny what he does, he learns, do something else. A bit like the beast in black & white!

forgotmysn 10y ago

pretty sure Rival Theory was working on something similar to that

_John-Titor_ 10y ago

There's a recent game called Hello Neighbour where the eponymous neighbor learns from the patterns you use to sneak into his house. The one video I saw of the game looked freaky. 10/10 would play.

squeaky-clean 10y ago

> Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy. > Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples? It's a really tough trade-off between fun and realism. It's cool that your units have personality for a bit "Oh, look, those 2 dwarves are fleeing from the troll fight early!", but not fun once they disobey your intentions "That fucking troll has 3 hit points left. If my 2 dwarves didn't run away, it would be dead! Now it's destroying my base!" One game I remember striking a good middle ground with this was the original Battle For Middle Earth. It changed animations, sounds and other aesthetic things based on AI, while the game remained the same mechanically. If your units ran up against a mountain troll, you would see the ones in front of it cowering and holding up shields, while units behind attack furiously, but every unit was really doing the same damage. Or if you left troops of Orcs just standing around too long, they'd break out into fights and kill other orcs, but the animation is done as a trick where as one orc model "dies", a new one spawns in the group. So you didn't really ever lose Orcs, even though there would be orc corpses strewn wherever you had been camping. I think the main reason though is just that AI doesn't really sell. Graphics, physics and voice actors sell. You can blame a failure on AI, but very rarely is success attributed to good AI in a game.

Karunamon 10y ago
Bad AI often winds up panned in a game's review scores, which have an impact on sales. For instance, bad pathfinding in an RTS game can make the game almost unplayable. Halo is just about legendary at this point for teammate AI that is worse than worthless [1]. Any one of these things can destroy immersion at a critical point. AI is to video games as IT is to a modern enterprise. If it's doing its job, nobody notices, but if it's bad, it sticks out like a sore thumb. [1]: https://www.youtube.com/watch?v=vdbg8h3d3uM https://www.youtube.com/watch?v=vdbg8h3d3uM
- organsnyder 10y ago
  Or, if it does its job really well, it gets accused of cheating. My former company had a backgammon game with a really good AI. We constantly got accused of cheating, even though the dice-roll algorithm was identical (it didn't have a parameter to specify who the roll was for). What people missed is that a good backgammon player will position themselves to benefit from a wider set of possible dice outcomes.

JabavuAdams 10y ago

Game logic has been parallelized since about 2005-2010.

BickNowstrom 10y ago

In ARMA opponent AI is a parameter. When you have a lot of players on one server, you turn down enemy AI so you don't bog down the entire server with pathfinding calculations.