Benefits of higher performance Console CPUs (Game AI) *SPAWN*

Averagejoe

Regular
If your program uses only small amount of memory, GPU should have ~6GB available to it.

IMHO, it's a lot more interesting to think what games will be able to do with >4GB of memory dedicated for a CPU..


For example.? Physics and things like that.?
 
Persistent world, AI actually interacting the world outside the players FoV, etc.

So, what is the limiting factor for what would be perceived as "good AI"?

Is it that the software algorithms for game AI are not advanced enough?
Is it that "character AI" does not have enough memory per character?
Is it that the code for good AI is compute heavy and would drag the system performance down?

Just curious...
 
Real AI doesn't exist in games, and won't for a long time. It takes much much more resources to produce something believable than scripting. Besides, people with chops in real AI programming can make absurd amounts of money building things that interact with the real world, why would they work for pittance in the game industry?

When we talk about AI in games, what we are talking about is script. The quality of the ai depends mostly on the quality of scriptwriting. Having more memory, and more "cpu-like" computing power. AI script tends to be mostly integer, not data parallel, very jumpy, and have a lot of random data access. This is a description of the kind of code xbox360/ps3 were absolutely awful at.

Simply having those jaguar cores in the machine is going to increase the total volume of script the next-gen consoles will be able to execute by a ridiculous margin. Having more memory allows for more game state and more script.
 
8 x Jaguar still sounds less ideal for AI than, say, 4 x Trinity. Let alone 4 x Core i5/i7 or 8 x Steamroller. Though not perhaps within the same power budget.
 
8 x Jaguar still sounds less ideal for AI than, say, 4 x Trinity. Let alone 4 x Core i5/i7 or 8 x Steamroller. Though not perhaps within the same power budget.

Well, you had the best AI of history running in a pentium 200. (Half-Life). And by the way Jaguar ( 4 cores ) multithreaded performance is similar or superior to an i3 at same speed.
 
Real AI doesn't exist in games, and won't for a long time...When we talk about AI in games, what we are talking about is script.
AI can be evaluated on many levels. At it's simplest, and the level at which humans operate most of the time, it's just decision making evaluating options and weighting them according to preference. You can also preload tactics, and even server-learn player behaviour to create complex AI behaviours. There's a whole thread on the value of AI, but I don't see either consoles lacking in decent game AI (in the other thread I've said I'm not a fan of true AI and value the dumbness of bots in games, otherwise I wouldn't stand a chance!).
 
Well, you had the best AI of history running in a pentium 200. (Half-Life).
Rose-colored glasses are great... ;) While HL2 had some pretty convincing AI for its time, it wasn't really THAT great. It depended entirely on navpoints and pre-set cover spots, enemies had infinite ammo, grenades and so on, and they were omniscient, always knowing the instant you lobbed a grenade (even if they couldn't see it!), as well as your exact location.

I don't think that qualifies as "the best AI of history"...
 
8 x Jaguar still sounds less ideal for AI than, say, 4 x Trinity. Let alone 4 x Core i5/i7 or 8 x Steamroller. Though not perhaps within the same power budget.

Since most AI boils down to search, the amount of memory available is far more important than the amount of processing power available. And since the AI in games is, for the moment, going to be restricted to fixed co-ordinate maps and very limited interactions that's not really going to change.

In the future if they start using VP systems, and make more use of NLP systems, then I can see the CPU power requirement increasing. But the day I need to dive behind the sofa to take cover or crawl on my belly towards the TV, or even climb into a real Cardboard Box(tm)... well that's the day I'll be swapping back to the PS3\2\1 as I'll be far too arthritic and creaky to be able to play!!
 
Since most AI boils down to search, the amount of memory available is far more important than the amount of processing power available. And since the AI in games is, for the moment, going to be restricted to fixed co-ordinate maps and very limited interactions that's not really going to change.
Well, things like vision cones, require quite a bit of ray tracing, and things like moving requires path finding. If it's a simple environment (no elevation changes, or partial obstructions), then these things are easily done, but as soon as you complicate the environment, these tasks can be quite computationally intensive, and the task scales linearly with how many NPC's are running the routines. You can pre-bake navigation paths if the environment is not dynamic, however, as soon as you put in dynamic elements, like boxes, or even destructible terrain, you need to do it real time.

The actual decision making of AI, though isn't very computationally complicated, a bunch of case statements really. Though, I don't make a living writing AI code...

I bet to get swarm and teamwork to well and realistically in a dynamic enviroment might be complicated, and perhaps relatively computationally expensive (and probably doesn't scale linearly to the number of NPCs).
 
So, what is the limiting factor for what would be perceived as "good AI"?

Is it that the software algorithms for game AI are not advanced enough?
Is it that "character AI" does not have enough memory per character?
Is it that the code for good AI is compute heavy and would drag the system performance down?

Just curious...

In today's highly connected world, they can use real human for NPCs, bosses and comrades.

They can also use server database for large scale AI. E.g., companies like Google have a huge database about their users to appeal to them.

I suspect the tricky part is the presentation of intelligence. e.g., AI should be balanced and fun, not overly difficult/easy. They also need to create a large library of info and animation to illustrate its rich behavior.
 
Well, things like vision cones, require quite a bit of ray tracing, and things like moving requires path finding. If it's a simple environment (no elevation changes, or partial obstructions), then these things are easily done, but as soon as you complicate the environment, these tasks can be quite computationally intensive, and the task scales linearly with how many NPC's are running the routines. You can pre-bake navigation paths if the environment is not dynamic, however, as soon as you put in dynamic elements, like boxes, or even destructible terrain, you need to do it real time.

The actual decision making of AI, though isn't very computationally complicated, a bunch of case statements really. Though, I don't make a living writing AI code...

I bet to get swarm and teamwork to well and realistically in a dynamic enviroment might be complicated, and perhaps relatively computationally expensive (and probably doesn't scale linearly to the number of NPCs).

For a fully dynamic AI, I was hoping LBP or PS Home would go there.

Let us construct our own Sackbots to compete and fight in assorted user levels. This should be up LBP's alley.

The most interesting thing about Home is when other players are engaged in their own mini-games, other players can still see their in-game animations and items (to a certain extent), kinda like an MMO where everyone live in a unified yet partitioned world. Hmm... I guess like SimCity ?
 
Let us construct our own Sackbots to compete and fight in assorted user levels. This should be up LBP's alley.

Been done several times over the years in various forms.
Problem is like all user generated content, quality is all over the place and 99% of it is crap, so <1% of your user base dominate, which makes it no fun for the majority of your players.

This is off topic, but It's not that Game AI ( and I use the term to describe enemy behavior rather than sim like behavior) is simple because that's all that can be done, it's simple because it's what provides the experience most game designers are trying to create. They want to control tempo, and pace in a level, and you can't do it with a bunch of autonomous AI's running around.

There is a computation cost in terms of evaluating the current world view of the AI line of sight, tracing possible paths to a target etc.

When most people describe AI's being dumb, it's usually referring to some un covered edge case, running into a wall standing stupidly and letting you shoot it, there just really hard to find and fix regardless of the actual decision making mechanism, often it's just bad data in the level, or undersampling of cast rays, the AI's world view is incorrect, it makes a decision based on it and does what appears to be incredibly stupid.

Having said that for anyone whose ever done a raid in WOW or any other MMO with an inexperienced group, there are plenty of Real Intelligences that will stand in the fire until the healer runs out of Mana then complain about the shitty healing job.
 
Well then there's one thing that AI should do more to be more convincing: complain! ;)
 
Rose-colored glasses are great... ;) While HL2 had some pretty convincing AI for its time, it wasn't really THAT great. It depended entirely on navpoints and pre-set cover spots, enemies had infinite ammo, grenades and so on, and they were omniscient, always knowing the instant you lobbed a grenade (even if they couldn't see it!), as well as your exact location.

I don't think that qualifies as "the best AI of history"...
HL 1.
 
Been done several times over the years in various forms.
Problem is like all user generated content, quality is all over the place and 99% of it is crap, so <1% of your user base dominate, which makes it no fun for the majority of your players.

Perhaps the construction tools and run-time are too low level.

The designers may plan for other built-in levels where dumb bots become fun to watch/play, instead of always pitting them against each other. Players should be able to reuse design parts too. They can also make the bots take input from Internet such as Google Search Trends, and assorted user votes/ratings/comments.


Well then there's one thing that AI should do more to be more convincing: complain! ;)

They can also steal/replay moves by real players. :)
 
If NPC AI required that much computation, it could be done in the cloud.

I'm still impressed by the experience of playing any of the Unreal Tournament games against bots, though.
 
Back
Top