Jump to content

  • Log In with Google      Sign In   
  • Create Account

Interested in a FREE copy of HTML5 game maker Construct 2?

We'll be giving away three Personal Edition licences in next Tuesday's GDNet Direct email newsletter!

Sign up from the right-hand sidebar on our homepage and read Tuesday's newsletter for details!


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

#41 Cessen   Members   -  Reputation: 122

Like
Likes
Like

Posted 12 May 2000 - 10:16 AM

We already have AI, it''s used in games all the time.
If what you are talking about it real intelligence, then call it that! Call it "real intelligence". AI is merely the artificial simulation of intelligence, which can be a mere calculator, since it can give you intelligent answers to given problems.
It seems that after the Matrix, that single line about us "giving birth to AI" has made people think that the AI in games and such right now must not be "real" AI. When I first saw that movie, and that line was spoken, I thought to myself "what the hell are they talking about? We already have AI."
Don''t get me wrong, I love that movie. All I''m saying is that the term "AI" is being mis-used.

Sponsor:

#42 NickGA   Members   -  Reputation: 122

Like
Likes
Like

Posted 12 May 2000 - 11:46 AM

Good point Cessan. I didn''t catch that in the movie. I was still trying to figure out why Morpheus hasn''t called me yet. I guess I better get a cellphone.

#43 Anonymous Poster_Anonymous Poster_*   Guests   -  Reputation:

Likes

Posted 14 May 2000 - 09:32 AM

I''ve read the whole thread, and I''d like to ask a few questions (not necessarily phrased as such, but they are) and make a few (possibly invalid) points:

First, why in the world would the earth represent the only kind of artificial intelligence possible? The entire evolutionary history of life on earth is based on a very small number of original forms. It''s not even necessarily true that DNA is the only encoding scheme for life itself, so the brain may or may not represent the only intelligence hierarchy. We''re subject to the environment, and most good ideas (ie neural nets et al.) still come from the natural world.

Second, there is the question of how accurately the human brain maps to a minimalist view of intelligence. It''s based on a set of layers developed over the course of evolution and it''s a really nice machine, but so are most SUVs. It''s just that power locks and power doors and their intellectual equivalents aren''t necessarily required in terms of pure function. On the other hand, the brain pulls some neat tricks outside the realm of chemical signals using a kind of structure called (I believe) a microtubule. The basic thing about this guy is that it brings quantum theory into view inside the brain. As anyone who''s studied NNs knows, the basic hierarchy of nets is essentially digital. Fuzzy nets may or may not get around that problem. But if we *could* design nets to be intelligent, they''d likely be a LOT smaller than the brain, probably orders of magnitude less expensive - which is still a lot of power, but it certainly helps!

Hmm. That moved away from games a bit too much, so let me justify: In the context of a game, it''s necessary to do one of two things - spend lots of time building a fast AI machine with grueling work every time or build one fast NN hierarchy once and then train it each time, spending lots of computer time in-game to make enemies "smart". But there''s no reason that a smart interpretation of output vectors can''t at least reduce the amount of work that has to be done - using, for example, a single hierarchy to simulate an entire class of soldiers, or even limiting NNs to COs, if you will, who then disseminate "smart" instructions to the dumber brute-force or traditional-AI forces. Not appropriate to all genres, certainly, but to enough to make things interesting, at least.

Last, but not least, there is the question of how well a NN captures the function of the brain. Fortunately, here there is a really good guideline in Creatures, which did a lot of neat stuff, particularly in the inclusion of "modifiers" which changed the effects of a variety of stuff in the net. Don''t remember the mag (New Scientist, maybe), but one of them had an interview with that dude what programmed it. Creatures, of course, had the luxury of CYCLES GALORE, but still. The brains there could learn to speak and eat and feel "scared" and all sorts of neat stuff. Recognizing "scared" as an output vector is a task for the designer: what should make a unit scared in the first place. Once that''s determined, the output is mapped. The class isn''t as clear as one might like, and multiple unique outputs might have to be considered equivalent, but the possibility does seem to exist.

Anyhoo, that''s my seven cents,
mikey

#44 Mantic   Members   -  Reputation: 122

Like
Likes
Like

Posted 14 May 2000 - 03:24 PM

A lot of very good young programmers have the habit of thinking in terms of ultimate simulation -- complex experimental applications of code to reproduce everything from the mechanical pysics of a muscle to the most minute reasoning of an artificial intelligence in it''s environment. This is similar to the writer whose every short story becomes War and Peace or the artist whose every sketch aspires to the Cistine Chapel.

This is something we, as creative talents, must learn to overcome in the interest of productivity. Think of it as optimization -- for a game the only concerns are those aspects which have some bearing upon the tasks and goals required to play. Or for pawns along the wayside, only the tasks and goals needed for them to achieve their purpose in the player''s experience. It is almost certain that leading computer games will never utilize perfect neural simulation as other areas will continually demand their own portion of the available space and processing power -- rendering ever more expansive and detailed worlds, endowing those worlds with ever more immersive aural depth, supporting broader and broader variety of inhabitants.

It is my firm belief that the ''virtual human'' will and should remain a denizen of the virtual laboratory, where the platform is always as powerful and specialized as it can be and the environment always geared to simulation without the restraint of game rules or performace.

#45 bogdanontanu   Members   -  Reputation: 122

Like
Likes
Like

Posted 16 May 2000 - 03:00 PM

Now maybe i should start a new thread....because this gets hard to read from start to end....

Hmm...Anyway i''d like to make a synopsis one again:

1.I stated that NN are/or can be almost exact copy of our brain an more important they can evolve (is a certain level of hardware is given) like any Real Inteligent ppl from this board

2.Some ppl douted that emotions can be "understood" by NN
others douted they can even work

3.Some exaples provided by other ppl that replyed here show that they work very well and more i stated again methods to evolve and DREAM and be creative....now about emotions i have to agree with another reply post witch explais that emotions are just "pattern of behaveiour" and so any NN will do them if proper trained.

4. Ppl are still afraid of this ultimate understanding of our "robotics" hardware....well some will always be

5. Some real problems where raised:

5.a Is today hardware capable to do a minimal NN to be human like inteligent...and so generate our first human being?
My Answer: is YES we only neeed some millions neurons up to a billion to understand the whole UNIVERSE and our hardware is now capable to provide that at a resonable dimension.
Downfall: we still have problems with interconections between those neurons....wires are the most expensive part in an artificial brain as in our brain (myeline isolations take almost 90% of our brain...)

5.b can we train a NN to "instant" have same IQ as a humman?
Answer: Not yet as i know it....(military ppl may know better). Problem is that it really takes time to train any beeing to be intelligent ... but if u train it long enought it WILL....

5.c can we copy NN once trained?
Answer: yes this is very easy...just make the same config and preset levels of trigger same as those of "trained" NN...and ooops u have an identicaly person created...better than clones i am sure....

5.d Some ppl have complaints about complex chemicals involved in brain functions...that we cant reproduce/understand yet...
Answer: Well do you think u know all phisics in ur PC microProcessor? (i think you dont even know how it works beyond basics) but you see it only sends electrical signals from one gate to another.. u can replicate it chemicaly or mechanics if you so like (only to make it huge) but it will work the same because chemical subst and electrical signals are only support for information they are not of the esence here......

6.Some ppl pointed out ...as i did also ... that there is a little more improvements in our brain (like tunneling over speed of light mybe...and bypass connections) but i think this are also not of the essence....NN is the main ideea here...and if it works just fine to make a humman being i think we can let improvements for later...

7. Some ppl think that "simple connections" cant be inteligent....lol...matbe they dont know how theit brain works..let me explain it a little:


General NN operations /trainning
=================================

A NN i maked by multiple layers of simple neurons
there are at least 3 layers (but usually many more)

First layer takes inputs from outer world (like pixels from video cameras) data form microphone, sensors for tactile inputs etc

Second layer makes connections each input to each output of the first layer this is the actuall brain so there may be many may more layers here

Third layer takes inputs form second layer and sends output to action making organs like voice or movement or other execution chains

Now how the hell does it work?

Loke this
A. present imputs with relevant data: like a cat to the video camera (you should start with circles or rectangles but we dont have time for that)

B.signals from inputs wil "fire" if certain (even random) levels are meet.Secondear levels will also fire some signals and so will third level (observe that all "fiereing" sequence depends on "levels") so some actions ... stupid at the beginning like the hummman will cry or even make "pipi" am himself will happen...

C.You the "mature person" know that the NN should say "this is a cat"... and dont like "pee on himself" do you
so tell the (actually do smthing drastic like yell at that thing) NN that he had done wrong (maybe even have Right or Wrong button installed)

D. NN knows she/he did wrong... so downgrade all levels that fired early and upgrade other (maybe even random) links levels...so next time you will get another response...
do so until you will get the proper response...(it may tahe a while but it will...u can even help a little in the beginning )

now this help is easy to do if this is the only thing she knows (just push the good damn "cat" link... ) but if more things are inside you have to NOT destroy the other good answers from before training....
this generatea a system of equations .....that you can resolve...lol mathematicians try this way...
or you can just push up random levels but from those involved in early lernning...because all of our world has a linked logic in it dose it not?

Well if you follow until here and you are not loosed... then u get the big picture ... dont you...?

I is Simple....and its in your brain and a machine ca do it the same way.... we may or may not have the time to create such a beeing yet....but we understand now that it can be done....very soon if not allready...

Nice eh...




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