Jump to content

  • Log In with Google      Sign In   
  • Create Account


What I think about neural nets now to do with games


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
10 replies to this topic

#1 rouncer   Members   -  Reputation: 285

Like
-1Likes
Like

Posted 18 February 2014 - 05:19 PM

Youve seen backprop defeat people with their own tactics.  done.
Youve seen creatures learning how to walk, even like a person.  done.
 
both cool things, but what else can you do?
 
TRIPPY NET GAMES
 
what people dont realize, is nets can control a video screen too, and thats what i never saw, i did never know.
maybe because they require the horsepower only a modern video card can give you... but back in the old days,
you could have done it on a 16x16 binary screen, instead... and that has possibly some joy attached.
 
so basicly, you can record a game with them, seems kinda pointless... cause why didnt you just play the game? 
 
and its more pointless when i tell you theres only a finite amount of different screens you can store... 
as if you simply recorded it with a screen recorder and each screen is a state... the naive implementation
is a single cell per frame but ill explain later how you can increase your state size theoretically.
 
about all you got over a screen record, is a method of compression, and screen matching powers... but this
is the fun bit about it.
 
but now what can you do?
 
a) you can control the game inside the record, by matching screens to the fresh controls, so you can play it
   out of the record.  so thats the beginning.
 
b) it can be ai... the computer can pretend hes you if he learns off your play, and he biases the screen
   to be the closest matched state, and comes up with your controls.
 
c) you can even hack the screen, and itll bring up the most similar screen in the record, so you can hand
   around a little man, and maybe a bird, and then itll start playing where you plotted it in.
 
d) noise can be introduced to some parts of the input to sway the game randomly around the place, and make 
   it more exciting.
 
e) you can search for different areas of the game by giving it labels, in a special label area it learnt
   whilst it was recording
 
f) with a bigger network with decent enough res and brightness you can teach it real video, and then if you 
   label the frames with controls and states, youll be able to play a game that looks real.
 
INCREASING STATE SIZE
 
well, the first thing you can do, is make your cell count a 4096x4096 texture, and thatll boost you up
to 16 million frames right then and there.  itll take you ages to fill that. To reduce synapses you can 
introducea a small hierarchy (making it a deep net) thatll group the screen more gradually, then if there 
were 20 active cells in the second to last region, the whole states are now only 20 synapses each, for example.
 
theres a heaps more complex way to do it, and its by using a distributed representation to restore whole
states.  its confusing, because the cells have to work as a team to some level of confidence produce
the parts that make a whole image, instead of garbage, which is what youd get otherwise.   
But then you could have possibly a billion states, and youd never ever fill it!
 
And not to forget, if you store animations in your cells, they have to predict the animation to get them
to activate straight away when you need them, then youll naturally develop alot more states also.
This will have to be used in combination with the distributed representations, and its very tricky, getting
it right.
 
 
CONCLUSION
 
the sad thing is, its not intelligence, and it must be 100% supervised everything it "knows" cause it
has to record it.   youll have to show it everything explicitly, so even if you had the states, making
the game that looked real would be impossible cause there would never be enough distinct video, for it
to fill it states up much at all.
 
The "real sim" would be a robotics venture, where you recorded the steering wheel with a camera on the
bonnet, and they would drive and drive and drive with the video autolabelled, but it would still take
god knows how many years to finish an explicit state set which could actually be interactive... if the
net was made professionally, distributed and predictive they probably never even hit capacity.
 
 


Sponsor:

#2 ApochPiQ   Moderators   -  Reputation: 14664

Like
1Likes
Like

Posted 19 February 2014 - 12:34 PM

What does this have to do with neural nets? And what evidence do you have that this is even possible, and not just some vague speculation?



#3 rouncer   Members   -  Reputation: 285

Like
-4Likes
Like

Posted 19 February 2014 - 01:27 PM

god, its simple, but its a whole lot of thinking thats not commonplace among programmers.

 

all its doing is recording bits of the screen into each cell, and the connections have a simple nn weight, which just multiplies a binary number, to a normalized scalar "confidence"  and then if you look down a little in the forum, youll see my initial work here, ive got a post there, its not a bogus demonstration,  but im a bit of a dope, when it comes to commentaries, i do not sound intelligent. smile.png

 

so more proof of the validity is in the post a litte down a bit, by rouncer, me.

 

 

trust me, the problem is, and its self defeating the purpose of doing this is->

 

explicit whole frames/states only  - as if it was just a screen recorder.    the cheat mode is using the combinations of the cell fires to represent whole states, as aposed to just sticking every whole state into an individual cell (which will give you very little states, unless there was a lot of cells, 4096x4096 - 16 million...) but im working on that right now, because its much better, way more store, but i dont think you can get whole states retrievable in full (from a distributed representation) like that easily, or actually capturing them all being possible at all...   cause magic isnt real.

 

but a neural network as i see it, is primarily a compressor and search engine.

 

 

it will be an awesome invention, if i manage to get to playback of 66 90 minute films out of a single plane (which may become a volumetric cube) of cells.  smile.png

its a new way to do video compression, but theres issues with it,  like it would only play on similarity to the last frame, and you can imagine how kinda wierd that would be. heheh


Edited by rouncer, 19 February 2014 - 01:30 PM.


#4 ApochPiQ   Moderators   -  Reputation: 14664

Like
1Likes
Like

Posted 19 February 2014 - 09:00 PM

How much actual information theory do you know? Compression is about eliminating redundancy, it isn't unbounded and it doesn't come for free. There is a hard limit on how much information you can encode into a single network without substantial loss of signal. What makes you think you can circumvent that?

#5 Álvaro   Crossbones+   -  Reputation: 12450

Like
0Likes
Like

Posted 19 February 2014 - 09:13 PM

Shouldn't we just ignore rouncer's ideas until he can produce one interesting tech demo? If the guy were simply an idiot, I would have some sympathy for him and I would try to help him with his misconceptions, the way ApochPiQ is doing.

However, every so often he posts things like this:

god, its simple, but its a whole lot of thinking thats not commonplace among programmers.


So not only is he an idiot, he is also an asshole with a superiority complex. I am soooo done with this guy.

#6 IADaveMark   Moderators   -  Reputation: 2401

Like
0Likes
Like

Posted 20 February 2014 - 12:25 AM

Whoa... I'm glad you guys dealt with this cat first.


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!"

#7 rouncer   Members   -  Reputation: 285

Like
0Likes
Like

Posted 20 February 2014 - 03:30 AM

look, if you just look a few posts down you can see my tech demos.

 

ok, i just told you! the states dont come for free, in the naive implementation its a whole frame per cell, so you dont get many frames.

 

but the theory goes, you can daisy chain the distributed representation, and then you get state loss over a representation without recording it, but you do gain a few states over a whole frame store per cell.

 

see the distributed representation, is a lot of parts, that make a whole picture together, that the reuse of gives you an exponential amount of states, and all im doing is getting the most out of it possible. thats it.

 

and isnt that what video compression is, more frames for less ram?

 

the easy way to think about it, is im not storing still things twice, but it even works for the animating bits. ;)

 

 

theres more, and i think im being ignored insubstantially.

 

 

i guess the nice ones will think, well, maybe hell come out with the fantabulous video compression, but its not really prooven by this ignorant little post, but thats why i dont blame you guys.

 

im sorry im just too dumb to know how lame i am,,,   ok i wont come back

 

like what do i want, a pat on the bum?


Edited by rouncer, 20 February 2014 - 03:58 AM.


#8 samoth   Crossbones+   -  Reputation: 4656

Like
2Likes
Like

Posted 20 February 2014 - 08:12 AM

How much actual information theory do you know? Compression is about eliminating redundancy, it isn't unbounded and it doesn't come for free. There is a hard limit on how much information you can encode into a single network without substantial loss of signal. What makes you think you can circumvent that?

What? I can compress any file of size N bytes (N>=2) to N/2 bytes using a O(1) algorithm. It's a super secret patent pending algorithm though, so I can't tell you details. Also I'm still working on the decompression part...

 

Seriously though, I can see how a neural network could be used to compress video at much higher rates than lossless or conventional lossy video compression. However, if that worked, it would likely work in a similar way our brain compresses memories, so the movie wouldn't be a lot fun to watch.

 

When you recall a movie that you've watched, or generally for every memory of yours, it is no problem whatsoever to completely fit it into your brain, in full detail. No matter how many films you watch in your life, and no matter how long you live. How does that work?

The brain only memorizes maybe 0.001% of the detail, and when you recall your memory at full detail, your brain fills in the missing information either with some plausible interpolation or with something completely unrelated that it has seen at some point, whichever fits best. You have no way of knowing the difference because it's your brain telling you, so it's a perfect "reconstruction". Even if you can't recall a particular scene right now, you very clearly remember the rest of the film, don't you.

Why do you think witnesses always tell such a lot of bullshit when they're questioned? biggrin.png

 

Now imagine that kind of thing for a movie that you are to watch on a TV screen...



#9 wodinoneeye   Members   -  Reputation: 721

Like
0Likes
Like

Posted 25 February 2014 - 01:23 AM

finite amount of different screens you can store.

 

 

reexamine the math of the problem you are proposing to solve

 

a smallish game screen (these days) 1024 x 768 = 786432 pixels

 

assume for simple example its just black and white  0/1 values per pixel (assume for later color values will be magnitude more)

 

what is 2 to the power of 786432  ???     (rough calc 3 decimal digits every 10 bits)

 

that is a number with about 235929 digits ...  your different 'screens' states

 

 

finite, yes.  horribly huge, yes.

 

 

 

you might want to think more along the lines of preprocessing your 'screens' with some radical data reduction like edge detection and then some kind of 'stroke' encoding scheme to crush the state inputs to your NN down significantly.(like by magnitude of many thousands)

 

Be aware though that the bigger a NN (of the type you seem to be refering to even reduced in the way Ive suggested) is the more unlikely it will successfully form.    Backtracking formation of the weights also requires many cyclings of the training data - not something you can do really  in anything like real time.

 

You also WONT be able to recreate the original raw input 'state' (the video screen) as all the (hopefully) extraneous detaiil would be crushed out of the internal net data - the best you could reconstruct would be (for the edge/stroke factoring example) crude vector lines.


Edited by wodinoneeye, 25 February 2014 - 01:28 AM.

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

#10 rouncer   Members   -  Reputation: 285

Like
0Likes
Like

Posted 01 March 2014 - 12:15 PM

it works, but noone believes it does, even you, so come back after youve thought for 3 years straight about it, and youll find it will work.



#11 FLeBlanc   Crossbones+   -  Reputation: 3085

Like
0Likes
Like

Posted 01 March 2014 - 12:49 PM

Shouldn't we just ignore rouncer's ideas until he can produce one interesting tech demo? If the guy were simply an idiot, I would have some sympathy for him and I would try to help him with his misconceptions, the way ApochPiQ is doing.

However, every so often he posts things like this:

god, its simple, but its a whole lot of thinking thats not commonplace among programmers.


So not only is he an idiot, he is also an asshole with a superiority complex. I am soooo done with this guy.


Easy there, killer. I'll admit he's said some blunt things, but they could as easily come from, say, not being a native speaker. But don't write him off completely; he has produced some interesting things in the past (eg, http://www.gamedev.net/topic/644452-distance-field-landscape-demo-with-attachment/ ) and he is at least working and developing ( http://www.gamedev.net/topic/652697-neural-network-motion-video/ ), rather than just talking out his ass.




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