Jump to content

  • Log In with Google      Sign In   
  • Create Account


Neural net is the ONLY real AI...


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
44 replies to this topic

#1 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 06 May 2000 - 11:11 AM

i have to tell u this: we all have a neural net in our brain...the most advanced neural net ever designed (to our actuall knowlege) but it is only a neural net that ca: -evolve in time by learnning from its mistakes and good things -have a starting pre-recorder (genetics) status as we dont want every creature to start all over again dont we? -understand all complicated concepts like --love --structure of universe --itself --etc And all is done only by small changes in its levels of triggering the "unitary neuron" It has 2 only 2 big problems: --1.its slow on sequencial computers...but it works --2.mathematicians dont qiut understand how it works and thus dont belive it. Actually there is no known theory to explain how to make the initial layout of neurons in the net. and some layouts will be faster (morte inteligent) than other ... but ALL wil become inteligent if more time is given.. Now on the other side Genetics is just the language used to thansmit programs for a newborn man (or neural net) so that his hardware knows waht to do to biuld the neural net and the supporting devices for it.... So genetics is just a programming language...maybe god''s programming language (so yes its very evolved) but no mre than that.... So IMHO if u guys wanna make real AI u have to study neural nets and experience with it... PS. there is also another problem: becuse they are so slow nobady had ever tryed this neural nets in a real time game.... maybe u can do it? take care not to make monsters....

Sponsor:

#2 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 06 May 2000 - 11:18 AM

ahhh i forgot:

Dont even think of leaveing it evolve on its own..
it will take bilions of years..like it took on Earth..

u have to train a neural net with semnificant facts of life
from where it can learn rights and wrongs...
maybe so it will take less than 1000 years...

but it can be done....if only somebody is not doing it now
then he will have the perfect AI machine ...and kill us all...as useless...

#3 Andreas Jonsson   Moderators   -  Reputation: 3265

Like
Likes
Like

Posted 06 May 2000 - 11:37 AM

I agree with that neural nets are the artificial intelligense that best resembles our own. Unfortunately humans have millions and millions of neurons which we cannot afford with todays computers. At most we could have about a 100 neurons per creature in our games, but that would not make them very smart. With a 100 neurons you can only recognize very few patterns, which is really the only thing neural nets can do: recognize patterns.


- WitchLord

#4 NickGA   Members   -  Reputation: 122

Like
Likes
Like

Posted 06 May 2000 - 11:57 AM

I agree, Neural Nets are very interesting and are based on the firing of neurons in the brain. However, the use of neural nets for games is still up for debate. Have you read some of the articles on AI here at gamedev? They offer some reasons why games have not utilized them. One problem, as you stated, is the structure is difficult to determine. Sometimes also, from experience, when you try to train a neural net, the movement of the weights go away from the optimal (for example bad learning rate). So as you can see, if you send out a game with a neural net, it is possible that a "smart" initial weighting for a monster, for instance, can turn into a "dumb" monster as the game progresses. They mention in some article here that they just reset the neural net if things go bad, but that gets rid of all the "learning".

Have you heard of Fuzzy Neural Nets? They might be more suitable since some games I hear use Fuzzy in their AI. A nice thing about this type of neural net is that you incorporate the uncertainty compensation of Fuzzy with the training ability of a neural net.

#5 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 06 May 2000 - 12:04 PM

Yessss patterns of our life
patterns of images
patterns of universe
patterns of behaveiour
patterns of love

All we are and will ever be are patterns...maybe of God
but still patterns (patterns=law if u think twice is it not?)

and yes we have a lots of neurons but that is at the grasp of modern computers and more at the grasp of modern hardware
as a neuron is only a trigger operational amplifier after all...

Not the number of neurons is the problem...
THE BIG PROBLEM is the number of connections each other with each other...(and secondary how to layout those connections)

We can make now days about milions of neurons on big hard drives but where to store the connections ...
1 million x 1 million = ohhh my god

As in humman brain where the connections and isolations for that connections (myeline) takes the most place in the brain
neurons beeing only a thin (grey) surface....

So maybe we need hardarw chips with neurons and connections inside (or outside to configure them)
i think this will happen soon if we want AI in our computers..... it has to spread like 3D acceleration today

Until then even 100 startegucally placed neurons (on top of a expert system) will make your game learn and adapt at least a little to humman tricks

And remember not to make monsters....please..



#6 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 06 May 2000 - 12:15 PM



where is that fuzzy logic article?

thanks...

#7 Chabba-Wabba   Members   -  Reputation: 122

Like
Likes
Like

Posted 07 May 2000 - 02:04 AM

Uhm... excuse me ?

The only real AI?

Actually, they''re a good approximation to a possible set of solutions.

A neural-net is not a real-time adaptive solution!

The ability to learn >WHILE< in use is something which will make it a true AI.

Yes, NNs are good models, but they''re by no means a solution to the AI problem.

I''d imagine that a hybrid solution would be the best AI solution given the current tech and available resources.

-Tim

#8 Anonymous Poster_Anonymous Poster_*   Guests   -  Reputation:

Likes

Posted 07 May 2000 - 02:13 AM

It has been tried before, the NuralBot for q2 for example and it sort of works.

#9 MikeD   Members   -  Reputation: 158

Like
Likes
Like

Posted 07 May 2000 - 02:32 AM

Of course neural networks can learn while working in real time. It''s just a case of fitting in the processing power, our own brains back propagate between examples whether that is between the last time and the next time you caught the ball in a game of catch or overnight as you sleep it doesn''t matter. The fact is, to learn takes time and processing involves your brain not being used give answers as the net is updated.

So, sure, you can put a NN in a game and update it in realtime as long as it''s useful to do so, it comes up with the right answers and it doesn''t take up too much processing power to do so.

Mike

#10 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 07 May 2000 - 12:08 PM

Of course that NN will always learn while in game ...(after all look at YOU.. )

there is no proble to that ...i anly said before that u have to learni it(or he/she) a little at the beggining so it takes more intelligent actions from start...after all u dont want an ameoba to command your Real Time Strategy Game

And also u can have the NN only on top af the comand chain...let some simple expert system deal with move,attack tactics and stuff....let the NN do only the main strategy...
so you dont consume all processor power...

#11 PyroBoy   Members   -  Reputation: 122

Like
Likes
Like

Posted 07 May 2000 - 01:23 PM

Does anyone know of a good Neural Net programming tutorial? I''ve heard a lot about these things but I have no clue how they work internally or how to begin coding a simple one...

#12 NickGA   Members   -  Reputation: 122

Like
Likes
Like

Posted 07 May 2000 - 04:03 PM

Pyroboy,

Here is a link to something under programming--artifical intelligence I found on gamedev. It should get you started.

http://gamedev.net/reference/docs/refarticlelistuser.asp?catid=18

Nick

#13 MadKeithV   Moderators   -  Reputation: 971

Like
Likes
Like

Posted 07 May 2000 - 09:16 PM

www.gameai.com
Best site on Artifial Intelligence to get started.
And about neural nets - they aren''t half as useful as you think Bogdanontanu. Thousands of hours of research have gone into them, and the only things they can be applied to reliably is pattern recognition. It seems the human brain itself is slightly more complex than the structure of a neural net.


#pragma DWIM // Do What I Mean!
~ Mad Keith ~
**I use Software Mode**


#14 NickGA   Members   -  Reputation: 122

Like
Likes
Like

Posted 07 May 2000 - 09:24 PM

Hmm.. only pattern recognition reliably? I''m pretty sure they could be reliably be used in control systems as well (if we get a good function mapping). I can''t remember which school off hand, but it''s somewhere in Texas where they are working with critic trained neural nets. I thought they were having some success there. Have you heard anything of this MadKeith?

#15 Anonymous Poster_Anonymous Poster_*   Guests   -  Reputation:

Likes

Posted 07 May 2000 - 09:42 PM

Well, I''ve done a bit of research on nns. Not much but enough to see how they work in the brain. And what I devised was a system optimised for games and other computer based applications. It basically reduces the number of neurons and links and also speeds up the learning process.
What I did was to replace the neurons with more application specific commands (like "Is Enemy Nearby", "Health Level" and "Move Forward") and hardcoded those decisions and actions. by doing this, the chances of bad learning are greatly reduced allowing for quicker learning.
The dendrites and axons (links between neurons) are called links and each one has a score out of 100. Whenever there is a success (Say the computer takes a knight in chess), the links involved recieve extra points. If something bad happens (Checkmate to computer) then the links will lose points. If the points for a link drop below the threshhold value then the link is destroyed and a random link is placed with a neutral score.
This system works quite well for many applications although it may help to put in a learning mode so you can evaluate it''s successes and failures.
The problems with this method is that a lot of the ''human'' aspect is lost, unless you specifically remember it. For example, if you consistently attack player 2 who is controlled by a neural net and it finally gets its chance to kill either you or player 3, it may still have a 50/50 chance of attacking either even though a human would have disrespect for you and be wanting vengence

my $0.02

#16 MikeD   Members   -  Reputation: 158

Like
Likes
Like

Posted 08 May 2000 - 12:45 AM

To MadKeithV,

Name one thing intelligent life does that isn''t based in pattern recognition.

Mike

#17 MadKeithV   Moderators   -  Reputation: 971

Like
Likes
Like

Posted 08 May 2000 - 12:48 AM

Being creative.


#pragma DWIM // Do What I Mean!
~ Mad Keith ~
**I use Software Mode**


#18 PyroBoy   Members   -  Reputation: 122

Like
Likes
Like

Posted 08 May 2000 - 06:06 AM

Thanks for the links :-)

I saw something about some guy in the states that uses hardware neural nets (a bunch of transistors) in little robots that look like bugs. These things can walk around, compete with each other for light (the solar powered ones) and some can even pray on other bugs by stealing power from their batterys. They have about the same intelligence as real insects, ie- not a whole lot, but they don''t contain any software at all. They weren''t explicitly programmed to do any of these things. Kinda neat.

#19 MikeD   Members   -  Reputation: 158

Like
Likes
Like

Posted 08 May 2000 - 08:05 AM

How do you determine whether anything you do creatively is good or bad? How do you decide what direction to head for with your creativity? Past experience. Mental adaption to previous stimuli. Pattern recognition.

Mike

#20 Buster   Members   -  Reputation: 100

Like
Likes
Like

Posted 08 May 2000 - 08:06 AM

Yes, but can we design a neural net that doesn''t lose nodes when we take a bong hit?




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