Jump to content

  • Log In with Google      Sign In   
  • Create Account

AI Creation


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
22 replies to this topic

#21 wodinoneeye   Members   -  Reputation: 877

Like
0Likes
Like

Posted 10 October 2012 - 08:09 AM

Wod, have you ever tried to use AI that learns on it's own, even within very limited parameters they will usually fail spectacularly or in weird uncontrollable ways.

If you then add any tweaking you may as well have just done your own logic to get the behaviour you want much faster and with more control.

I have heard stories of completely learning AI in military, such as one that learnt to detect hidden tanks, they taught it using millions of pictures. Then when they took it out into the field to test it, it gave completely messed up readings: tanks were everywhere! After many programmers poured through many logs they found that the AI had actually picked up a specific shape in clouds that it thought was a tank, basically a waste of money.

You have a (completely) learning AI and vs it against my fully scripted AI and mine will beat yours the first million times while also acting more like an adult human.



Yes I did NN (not for games). And look above and see that much of the process I talk about is heavily human guided. The AI development programs/processes I mention are more accumulation tools than anything (alot of the grunt work is done by the program, making better use of the users time -- the chokepoint...)

Its not so much 'tweaking' as telling the machine yes thats right (your logic is sound) or no its wrong for that case - revise you logic like this for that special case (which accumulates as you go on).

Nowhere above do I even mention Game AI that 'learns on its own'.

I did mention recording and factoring from observing a real player playing the game - but thats test data gathering and a AI developer is still needed to shape the logic using the data as examples. Even just factoring the situation properly (so it can be turned into the 'if' part of behavioral logic) is a hard task - theres a whole area of AI 'cognitive' which is about doing that.


-----------

I actually professionally was evaluating some commercial face recognition software and was amused to find that it was detecting human faces in odd reflections and patterns on texture painted walls and then wouldnt read my face right in front of the camera because of the glasses I wear or if I put a hardhat on.

Edited by wodinoneeye, 10 October 2012 - 08:17 AM.

--------------------------------------------Ratings are Opinion, not Fact

Sponsor:

#22 Olof Hedman   Crossbones+   -  Reputation: 2949

Like
2Likes
Like

Posted 13 November 2012 - 07:15 AM

I've heard of some game that used a NN, but I don't recall wich now.

But for most games, you get the behavior you want much faster and easier with other techniques, and thats the point.
You can get very challenging and hard-to-predict agents with just an FSM and some utility-functions, you can even get it to learn if you want...

You seem a bit confused on what AI is though... If an FSM does the exact same job as a pretrained NN, why would the former be any less "intelligent" then the latter?

NNs are useful for a rather particular subfield of AI, and isn't really the magic "simulated brain" that many seem to assume when they first hear about them...

Edited by Olof Hedman, 13 November 2012 - 07:19 AM.


#23 IADaveMark   Moderators   -  Reputation: 2531

Like
2Likes
Like

Posted 13 November 2012 - 08:36 AM

I've heard of some game that used a NN, but I don't recall wich now.


You are probably referring to Supreme Commander 2 that used an NN specifically for the tactical AI in skirmish mode. It was off-line learning that, in the end, amounted to a shortcut for tuning a utility-based system to get the parameters right. This was done post-ship on the fly because their original shipped AI was dogshit. They brought in Michael Robbins (from their modding community!) to fix it... and he was in a bit of a hurry, naturally. You could have gotten the exact same results via a hand-crafted utility-based system for tactical decision-processing.

But for most games, you get the behavior you want much faster and easier with other techniques, and thats the point.
You can get very challenging and hard-to-predict agents with just an FSM and some utility-functions, you can even get it to learn if you want...


This. Many people dismiss simpler techniques as being incapable of doing clever things simply because they don't do clever things. What they don't realize is that seemingly dense AI agents are often a design decision rather than fallout of a less-than-capable AI system.

You seem a bit confused on what AI is though... If an FSM does the exact same job as a pretrained NN, why would the former be any less "intelligent" then the latter?


Covered this earlier... the simply use of the term "artificial intelligence" sweeps up academic-minded people in the net who then scoff that the "AI" we use in games really isn't "AI" as understood in their world.

Get over it folks. Posted Image

NNs are useful for a rather particular subfield of AI, and isn't really the magic "simulated brain" that many seem to assume when they first hear about them...


And that's the really funny part... NNs are actually looked on as a bit passe in the academic world as well. If they had a less sexy name, people would have long since dismissed them.
Dave Mark - President and Lead Designer of Intrinsic Algorithm LLC

Professional consultant on game AI, mathematical modeling, simulation modeling
Co-advisor of the GDC AI Summit
Co-founder of the AI Game Programmers Guild
Author of the book, Behavioral Mathematics for Game AI

Blogs I write:
IA News - What's happening at IA | IA on AI - AI news and notes | Post-Play'em - Observations on AI of games I play

"Reducing the world to mathematical equations!"




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS