By using this site, you agree to our Privacy Policy and our Terms of Use. Close

Forums - Gaming - Next-gen AI and Next-gen BS

vivster on 26 March 2014
Salnax said:
vivster said:
Salnax said:
So what I'm getting from this is, I was right in assuming there would be no great AI leap this gen?

That would be the pessimistic way to look at it.

It seems to be completely up to the developer to create a great AI. So maybe their motivation will be reignited by a new generation.


Considering how the emphasis on games that require little AI does not appear to be changing (online shooters, platformers with dumb enemies, etc), I doubt it.

Honestly, back in the day, I had one great hope for AI the 8th generation: Nintendogs 2, later named Nintendogs + cats. I really haven't had much hope for new advanced AI since the 3DS launch came and the virtual dogs were still virtual dogs. I mean, if the leading pet simulation series is not going to revolutionize AI, what would? Say what you want about pet sims, they're based around a single character's AI and responding to it.

I don't doubt that AI will improve here and there by increments. But it seems that for every tenfold improvement in graphics we get, AI gets a few IQ points.

Do you think, if VR hits it off, that would give devs more reason to program more lifelike AI?



If you demand respect or gratitude for your volunteer work, you're doing volunteering wrong.

Around the Network
theRepublic on 26 March 2014

Interesting article. I never really gave much thought to how much computing power the AI was taking in a particular game, and it makes sense that it is very little.

I would guess I am like most gamers. I only really notice the AI when it is doing something very stupid, or it feels like it is cheating. Pretty much the two scenarios covered in the article.



Switch Code: SW-7377-9189-3397 -- Nintendo Network ID: theRepublic -- Steam ID: theRepublic

Now Playing
Switch - Super Mario Maker 2 (2019)
3DS - Phoenix Wright: Ace Attorney (Trilogy) (2005/2014)
Mobile - Yugioh Duel Links (2017)
Mobile - Super Mario Run (2017)
PC - Borderlands 2 (2012)
PC - Deep Rock Galactic (2020)

theRepublic on 26 March 2014
Norris2k said:
Yeah, clearly that's BS, the guy doesn't even know what he's talking about.

I will make a digression on strategic IA (not theft one), I'm not even sure it's possible to have a good (strategic) AI.

The reason the cpu cost is low is that it's not even an AI, it's just a (hard to code and hard coded) collection of a few tricks and scripts relying a lot on cheats. You script a build order that will execute perfectly, add some build order rules (don't build to extend max cap if you are far from it), use cheats (know where the player is, cheat on costs, know player units), and set a few attack rules (for example send units when you have 10 units if player has less than 7). You will make it more complex, with added rules and build order, but it will never be enough to give focus, game and map understanding, long term strategy.

For example in Starcraft the AI will follow a predefined and optimal build order, and send units to you, directly to your base (no fog of war). The idea is to put you at a disadvantage that will make the battle harder (or even beat you), and the AI seems better. At some point, as the AI has no focus, no strategy, no understanding of what battle it can win or lose, and does not even have micro management. It just follow unrelated and suicidal directives, so you will overcome and prevail. With some experience, you will be able to manage it to the point you can handle 4 to 7 AI. 1 versus 4 is a lot more than 4 times more powerfull, it's something you can't achieve with players that have more than 10 hours of training.

Go beyond that seems extremely difficult. For example, the go game is very simple compared to a computer game. Only 1 unit, very simple rules easy to code, and a very limited and regular board : 19x19. Lots of people are trying to make an AI, and so far the results are not really impressive. It's not even real time.

Chess was solved by brute force, a solution based on search tree because the number of possibility (what you can do in one turn) is limited enough in term of possible moves at every step... to have a super computer from top 500... search for the next move for one unit... in 5 minutes !

Of course we are not talking about 'real' AI.  We are talking about video game AI.  As in, how the game plays with, against, or how it reacts to the player.  That was very clear in the article.



Switch Code: SW-7377-9189-3397 -- Nintendo Network ID: theRepublic -- Steam ID: theRepublic

Now Playing
Switch - Super Mario Maker 2 (2019)
3DS - Phoenix Wright: Ace Attorney (Trilogy) (2005/2014)
Mobile - Yugioh Duel Links (2017)
Mobile - Super Mario Run (2017)
PC - Borderlands 2 (2012)
PC - Deep Rock Galactic (2020)

Hynad on 26 March 2014
Norris2k said:
Yeah, clearly that's BS, the guy doesn't even know what he's talking about.

I will make a digression on strategic IA (not theft one), I'm not even sure it's possible to have a good (strategic) AI.

The reason the cpu cost is low is that it's not even an AI, it's just a (hard to code and hard coded) collection of a few tricks and scripts relying a lot on cheats. You script a build order that will execute perfectly, add some build order rules (don't build to extend max cap if you are far from it), use cheats (know where the player is, cheat on costs, know player units), and set a few attack rules (for example send units when you have 10 units if player has less than 7). You will make it more complex, with added rules and build order, but it will never be enough to give focus, game and map understanding, long term strategy.

For example in Starcraft the AI will follow a predefined and optimal build order, and send units to you, directly to your base (no fog of war). The idea is to put you at a disadvantage that will make the battle harder (or even beat you), and the AI seems better. At some point, as the AI has no focus, no strategy, no understanding of what battle it can win or lose, and does not even have micro management. It just follow unrelated and suicidal directives, so you will overcome and prevail. With some experience, you will be able to manage it to the point you can handle 4 to 7 AI. 1 versus 4 is a lot more than 4 times more powerfull, it's something you can't achieve with players that have more than 10 hours of training.

Go beyond that seems extremely difficult. For example, the go game is very simple compared to a computer game. Only 1 unit, very simple rules easy to code, and a very limited and regular board : 19x19. Lots of people are trying to make an AI, and so far the results are not really impressive. It's not even real time.

Chess was solved by brute force, a solution based on search tree because the number of possibility (what you can do in one turn) is limited enough in term of possible moves at every step... to have a super computer from top 500... search for the next move for one unit... in 5 minutes !


You, sir, win just because you mentioned my favourite board game of all: Go.

But I don't understand what you mean when you say the AI developed for this game aren't real time. I play on CGoban, on KGS Go server, against what we call bots (like GNUGo) and it's all real time. Sure, they're no very strong. The strongest I've heard of is around 2-3 Kyu level. They understand life and death pretty well. But they get overwhelmed quite easily if you use an ounce of strategy.



Norris2k on 26 March 2014
vivster said:
Norris2k said:
Yeah, clearly that's BS, the guy doesn't even know what he's talking about.

I will make a digression on strategic IA (not theft one), I'm not even sure it's possible to have a good (strategic) AI.

The reason the cpu cost is low is that it's not even an AI, it's just a (hard to code and hard coded) collection of a few tricks and scripts relying a lot on cheats. You script a build order that will execute perfectly, add some build order rules (don't build to extend max cap if you are far from it), use cheats (know where the player is, cheat on costs, know player units), and set a few attack rules (for example send units when you have 10 units if player has less than 7). You will make it more complex, with added rules and build order, but it will never be enough to give focus, game and map understanding, long term strategy.

For example in Starcraft the AI will follow a predefined and optimal build order, and send units to you, directly to your base (no fog of war). The idea is to put you at a disadvantage that will make the battle harder (or even beat you), and the AI seems better. At some point, as the AI has no focus, no strategy, no understanding of what battle it can win or lose, and does not even have micro management. It just follow unrelated and suicidal directives, so you will overcome and prevail. With some experience, you will be able to manage it to the point you can handle 4 to 7 AI. 1 versus 4 is a lot more than 4 times more powerfull, it's something you can't achieve with players that have more than 10 hours of training.

Go beyond that seems extremely difficult. For example, the go game is very simple compared to a computer game. Only 1 unit, very simple rules easy to code, and a very limited and regular board : 19x19. Lots of people are trying to make an AI, and so far the results are not really impressive. It's not even real time.

Chess was solved by brute force, a solution based on search tree because the number of possibility (what you can do in one turn) is limited enough in term of possible moves at every step... to have a super computer from top 500... search for the next move for one unit... in 5 minutes !

Usually Shamus is really insightful in his columns so I'll believe him over "random internet guy".

But you seem to have knowledge in this matter so please enlighten me about the other type of AI that is in shooters for example. Do you really think their AI up to now was really compute heavy? Are you saying it's more than a fraction compared to the rest of the process?

I mostly agree with Shamus, in fact, I think it's a misunderstanding (sorry for my english level). Basically I'm agreeing with him that AI take no CPU in games,and that the talk about thieft super-AI expanded thanks to new CPU is bullshit.

So, yes, AI for shooters are done the way he says and take negligeable amount of calculation, so there is room to improve it, because it's cheap check, and decision tree addition (that doesn't make it easy to make or to test).

What I tried to add to his talk is that even AI for strategic games are based on a similar collection of complex tricks and scripts, not an AI in the way most people would think it is. So, my digression is a doubt about if it would be achievable to have a real AI. And I think (imho) it's not, because a thinking opponent, if possible, takes years to code and a tremendous amount of calculation. But it's not about having a CPU twice or 10 time faster, it's about having a massive calculation power.



Around the Network
JoeTheBro on 26 March 2014

I pretty much agree with him.

The thing he was wrong about was the ipod gym analogy. Yes compared to other tasks AI calculations are minimal, but they still take valuable resources.


Check this link out: http://www.polygon.com/2014/3/17/5520290/battlefield-4-dragons-teeth-ballistic-shield
"We didn't have enough memory," Duffin said. "It added about a megabyte, but a megabyte on a 2006 console is a lot, so it didn't work out."

Both PS360 have about 512 MB minus their OS reserves, so with BF4 this essentially says the game runs at over 99.8% capacity as far as RAM. It's a similar situation with the whole console. Every little bit of power is used. Stronger machines can dedicate more resources to AI, even if it's a small change compared to everything else.



Links

Proclus on 26 March 2014

Very true article, I used to do a bit of programming, as far as shooters and RTS games are concerned you aren't going to take up a lot of CPU.

It's going to take that one brilliant dude/dudette to leapfrog AI a bit, under the current circumstances most developers don't really have a reason, financial or otherwise to make a serious effort at "next-gen" AI



Ssenkahdavic on 26 March 2014

Skynet is here.



 

  

The Only True Next Gen Game - Hero-U

Never underestimate the power of human stupidity.
Robert A. Heinlein

SvennoJ on 26 March 2014

AI used to take up quite a bit of processing power, a long time ago. In the later stages of Civ it could take up to a minute for the computer to finish it's turn on a 286. Path finding in baldur's gate was dumb as rocks, cause the search depth was set very low to save cpu cycles.

Some games are still limited by the exponential increase in path finding difficulty. The AI in Sim city 2013 works with a very simplified system which is responsible for the dumb traffic behavior. It's not that difficult to write an optimal distributed path finding solution, but the cpu cost increases exponentially with grid size and number of actors. Pretty much the same as trying to calculate the next move in Go.

AI so far is either a clever set of decision rules (cheap but predictable), or brute force determining the best move out of all current possibilities (very expensive or otherwise dumb) It's 2014 and Dark souls 2 enemies still have the same predictable behavior (intentional to keep it playable) but also plenty of exploits due to the pathing algorithm. We're still not at a stage where AI characters can move around naturally. (To be fair, neither do human controlled characters)



walsufnir on 26 March 2014

I didn't read the wall of text but to people interested in AI:

First: Don't confuse artificial intelligence with artificial life (often done in research)
Second: I don't know about videogame AI but if people are interested they should read articles about game trees (http://en.wikipedia.org/wiki/Extensive-form_game), alpha-beta-pruning, POP (partial order planning) (http://en.wikipedia.org/wiki/Partial-order_planning) and related articles.