| |
| |
| |
|
Comments:
<0> captain fourier heh
<1> hey all
<1> is there a Hangman game that is played by a computer based off of word/character frequency analysis/statistics and dictionary of words? (not necessarily for #AI, but close enough methinks)
<2> hi. not familiar with such a thing
<1> hmm ... well, suppose you are, are there any alternatives to frequency analysis of previous situations for selection of output (the letters to test and in what situation to try them) ?
<1> *and in what position to try them
<2> a dictionary?
<2> it's about asking for letters in a word and ultimately guessing it right?
<1> yeah, but actually I was hoping to disconnect the question from hangman
<1> I was thinking of general applications, as I would not like to be stuck to "frequency analysis" of situations if there is something better to consider
<2> what is 'general applications' ?
<1> a situation where software is used instead of human input in response to some other input/data?
<2> that is mighty general. probably so loose that it loses meaning
<1> alright, I suppose I will go sit in the corner pondering my question more
<2> you want to find a technique that will be the key part of a computer player
<2> that is essentially AI
<1> nah, not really, I was approaching this from a different direction: searching for patterns in a dataset and then making these patterns predictive models to work off of, but the only sort of way to make it 'predictive' seems to be through frequency analysis, which I am rather skeptical of
<2> the survey answer to that is that no such general technique is known and tested. for game theory more specific compromise solutions would often be applied. like IBM's Deep Blue where a specific solution is found for a specific problem
<2> but pondering is the right approach
<2> Einstein said he didn't possess any great intelligence. he just stayed with a problem longer than most
<2> a lesson that is likely underestimated by most
<2> think about what prediction means. what 'patterns' are. what intelligence is. examples often help
<2> 1,2,3,... up to more complex patterns like motion of a car from visual input etc
<1> the working definition of intelligence that I use is behavior generation with respect to some goal
<1> 'erm, sorry for that random message there
<2> sure but that will also need some refinement, naturally. for instance could a p***ive machine be intelligent?
<1> What?
<1> that does not seem relevant given my definition (behavior generation)
<2> your definition included behavior. does it really need behavior to be intelligent?
<2> would I not be intelligent if I sat still
<1> the point of using words is to have some framework to discuss things, I was merely telling you what I am using intelligence to refer to
<1> if you want to talk about things that do not do anything (no behavior) then be my guest O.o
<2> don't apologize
<1> heh
<2> I was merely showing you where your defintion fails
<1> ah, but you ask if you are intelligent if you just sat still
<2> think of HAL silently observing
<1> but you forget that you are in fact not sitting still
<1> there are many billions/trillions of components that are behaving to comprise your systems
<2> the key to greatness is to not let ego tie you down. with these things I point out you have two choices. pretend I'm wrong or grow with them
<1> what??
<1> I am talking with you, not pretending you are wrong ;)
<1> anyway, what particular problems are there with my definition?
<2> intelligence does not require behavior
<1> if you define intelligence as "the ability to generate relevant behavior with respect to some goal(s)" then yes it does, right?
<1> what do you suggest?
<2> do you understand that truth does not bow to your definition?
<1> again, what do you suggest for the definition of intelligence?
<2> while it might seem statistically very unlikely I am uniquely able in certain areas. you can choose to benefit from. I define intelligence in two variations. one static and one dynamic. static intelligence is the rigid ability of a engineered solution such as deep blue. a fixed skill. dynamic intelligence is the ability to learn new abilities. you can choose which you find more appealing
<2> from it*
<2> dynamic intelligence is the kind that starts from scratch (drooling baby) to fully competent
<2> static intel could be chess player (Deep Blue), walking droid (Honda), any kind of fixed prediction skill etc
<1> there is no conflict in our definitions, yours happens to use other terminology (problems, solutions, learning)
<2> mine does not require behavior. yours do
<2> think of the observer
<1> so you're telling me that 'learning' is not a behavior?
<1> and the same with 'solving' ?? /me is confused
<2> I see. you consider behavior as any activity even just internal computation
<1> sure, we could say that
<2> that is what you are saying :)
<2> a simplification of your definition is then simply 'goal achievement'
<2> which has the benefit of being true
<1> that is always useful heh'
<2> it's a good quality :)
<2> which do you prefer: designed to do a skill or designed to be able to learn any skill. (static vs dynamic intel)
<1> the second one is an instance of the first
<2> quite true
<2> that was quicker than the others :)
<2> ah well a joke. while it is true that the second could be said to be a special case of the first I think you know the intended distinction
<2> rigid deep blue vs drooling baby
<1> hmm, not entirely, for I have never seen a single skill completely isolated
<1> even adding two numbers is an accumulation of many systems
<2> how about Deep Blue? fairly discrete
<1> yeah in comparison certainly
<1> what do you know of neural networks (either biological or artificial) ?
<2> artificial. some. I am a proponent of them
<1> I see, have we any tools to highlight neuron firing patterns in our artificial systems?
<2> the key word in what I call dynamic intelligence is learning. indeed learning is fundamental to AI. machine learning, at which neural nets are uniquely qualified
<2> you could use a speed marker to highlight it. but I take it you mean something more specific by highlighting
<1> yes, I am ***uming that there are particular neural structures/pathways that develop over time
<1> (mind if we take this private?)
<2> why?
<1> hmm, not sure if I like the prospect of everything being logged
<1> I could live with it, though
<2> truth is a good thing
<3> at least i'm reading.
<2> 'Machine learning' by Mitchell has around 10 pages on neural nets. which is enough. the learning principle is quite simple and should be enlightening
<2> the book can be found on edonkey or in book stores
<1> I see, are there any resources that are incredibly complicated (while thorough)?
<1> (on neural networks)
<2> it's thorough enough. don't look for complexity where none exist
<2> unless you want to fail
<1> neural networks are complex; I have read a few articles that simplify them, but I find myself still not knowing what's going on
<2> NNs are complex not in their principle but how to apply it to real world situations. indeed its simplicity has fooled many scientists into thinking that it's a dead end
<2> that's a common mistake for an article. in trying to make it clearer they make it harder because nothing useful was conveyed to the reader
<2> only the false impression that it's impossible to understand
<2> 10 pages. not terribly complicated. the base concept of 'gradient descent' is very simple but it will take some time to fall into place for its context
<2> it's more or less y=bx finding b iteratively for a specific set of X and Y. little higher, little lower. there
<1> I see, sorry for my delaqy
<1> *delay
<2> gradient descent
<2> np
<2> come to think of it, I have the pages in question online: http://www.df-cad.dk/web/ai/[Mitchell]MachineLearning(page80-104-neuralnets).zip
<1> loading
<2> it might be moderately illegal but there is something about the fair use act allowing some pages to be copied for educational purpose
<2> when it comes to understanding how that seemingly irrelevant technique can be applied to reality you can look at the link in the topic
<1> what technique?
<2> Neural nets
<2> (backpropagation)
<1> let's say we have a data processing unit, running some software, and we have a motor system (like for Honda's robot). Is it possible/useful to attach a neural network from the software's output to the motor input/outputs ? Or would the entire 'brain' have to be a neural network taking feedback from vision-input components? DanF_DrC, ?
<2> the key skill is usually what you need the NN for
<2> it's not a gear box or catalyst
<1> but can the NN be used for very specific output control with some software saying the result was what was expected/predicted?
<2> the basic shape of a neural net is a function. to do a specific thing (output) in a specific situation (input). any situation where that applies you can use it
<2> f(x)
<2> it could be the entire system of just a part of it
<2> in the car example the input is an image of the road and the output is the steering direction
<2> input to output. a function in the math sense
<2> f(x) where x can be a vector (series of values)
<2> set
<2> NN is trained with a training dataset which is a set of correct examples of intput and corresponding output. the NN using backprop then adapts to that function. 'learns'
<2> input*
<2> it's a function approximator
<4> hi
Return to
#ai
or
Go to some related
logs:
#mysql
#sed
gentoo saslpasswd2 does not exist
libdb.so.2 fedora 5
#ai
shorewall configuraton
gentoo + fluxbox = font is blurred
varchar or tinytext
LBA-48 Centos
#linux
|
|