| |
| |
| |
|
Page: 1 2 3 4 5
Comments:
<0> though i think big changes are needed to make it work fast
<0> (it's theoretically bias optimal, but that means nothing)
<1> For complex problems I doubt it is practical
<0> _YKY_, it's possible, but TDD is a great methodology to reduce complexity
<0> it's like robot shaping, where you put your learning robot through a series of increasingly difficult trials to teach it.
<0> this would be shaping of a coding tool
<1> Hmmmm
<2> how are you going to formally write the tests
<0> I'm guessing the prototype will be python code, standard unittest
<0> (there are lots of examples out there, so i can test the system on common data)
<0> _YKY_, sceptical?
<1> Sometimes the solutions do not build up incrementally
<1> Yours work only for incremental problems
<0> _YKY_, no, it would work for any kind of problem
<0> but incremental would be faster
<0> and TDD by nature is incremental
<3> <aevin@ef> this sounds hard, but interesting.
<3> <aevin@ef> let me know how's it going along, AJC :)
<3> <aevin@ef> *it's
<1> For non-incremental problems the solutions to simple cases do not help
<0> aevin: I think i'm about ready to start implementing, i think i worked out the basic of the learning
<0> _YKY_, there are always examples that help "solve problems" generically... like oops, the system will meta-learn
<0> i'll go back to Christopher Alexander's original definition of patterns, and learn patterns for solving problems at all levels of the system implementation/development methodology.
<2> suppose you want to make program that does a fibonnaci sequence of n numbers
<2> how would that work
<2> it will just learn if( n==1) print 1
<2> if(n==2) print 1 1
<2> etc
<0> if it learns memoization, that's not bad. in that case, i give it a larger sequence of examples
<2> thats the point, it will just add more if lines
<0> then, the cost of implementing fibbonacci would be lower than using a table
<0> ttt-, well, fibbonaci is a bit of a silly example.
<0> but you could write a test that only gives certain examples, and checks against others
<0> or simply tell the system to not use memoization
<2> yeah but
<2> how does it figure out the fib. computation (the rule)
<0> it searches through a tree of possible programs
<2> it cant, apart from blind searching for it
<2> yeah
<2> making it intractable for almost all problems
<0> the searchspace here is (*, +, _input_, _recurse_)
<2> you have the search and the evaluation part figured out, but there's no actual AI in there
<0> huge numbers of AI systems are based on search
<2> but you can still add it
<0> but true, the challenge in the system is getting it to learn while it's searching, and search faster
<0> then allowing it to create generalizations after solving problems
<4> i haven't been paying attention. are you trying to do a GP implementation to learn the fib sequence?
<0> ttt-, OOPS managed to solve the hanoi problem with something like 31 discs, with recursion etc.
<0> *towers of*
<0> i think search space size is a problem, but i'm not worried about intractability!
<2> yeah because the search space is small for that problem
<2> if the number of lines of code is only 20, you have a small search space
<2> and you can brute search for the program
<0> brute force would not have solved this problem
<0> it required the meta-learning, checking ***umptions while solving, etc.
<2> i mean search with some heuristic
<0> what's the heuristic?
<2> i dont know
<0> exactly. that's the problem i'm trying to solve, automatically learning the heuristic so the searchspace is manageable
<1> You know Eurisko?
<0> no, do you have a url?
<1> An early AI program by Doug Lenat
<1> Automatically learns heuristics
<1> It was very old stuff though
<2> you can do only so much with search spaces and heuristisc
<0> ttt-, everything is a search space
<2> no it's not
<0> EVERYTHING!
<0> :-)
<2> if you want a program to code programs, why doesnt it steps like a human does
<2> learn about if and while loops
<0> though, granted, heuristics can only accelerate your search on a flat level
<0> aah, i was getting to that. that's where chunking comes in
<0> chunking is a form of hierarchical learning they use in SOAR. essentially, you're storing knowledge that resulted from heuristic search
<2> whats chunking
<0> though the challenge is how to apply it here.
<2> what kind of knowledge?
<0> OOPS already allows you to write such "hierarchical" solutions, you cohuld have actions for "implement-loop" etc. it's still a search.
<0> ttt-, SOAR uses production rules
<0> what OOPS doesn't do is induce the knowledge automatically though
<0> so it can't learn for loops from examples
<2> why dont you make an AI program first that understands simple programs
<2> that understands for(i=0...4) and for(i=5..6) equals for(i=0..6)
<2> and can reason with things like that
<0> it's a valid approach, though not one i want to take.
<0> i want the system to learn from solving problems only, goal driven.
<2> but if it doesnt know the simple things, how can it simplify programs?
<5> Lenat and Doug on Highlight list
<0> the risk of working on such "analysis" is that it's not necessarily useful in helping the system solve problems
<0> dmiles_afk, LOL
<0> ttt-, it does not simplify programs. its search is biased to programs with low complexity
<2> knowing "for(i=0...4) and for(i=5..6) equals for(i=0..6)" is so much easier than searching for a smaller equivalent program that p***es all the tests
<0> ttt-, i don't believe so.
<0> but we'll see :-)
<2> i really find this weird
<0> how so?
<2> why nobody is interested in making computers understand simple things
<5> AJC.. progress on alive?
<2> how can you except anything useful from a program that doesnt even know how to do " for(i=0...4) and for(i=5..6) equals for(i=0..6)"
<0> well, two comments. 1) lots of work is spent on understanding, but it's not necessarily tied to acting. 2) i don't model that understanding explicitly; but i expect it to develop useful knowledge and abstractions nonetheless.
<2> unless you got infinite computing power
<0> ttt-, if you generate your solutions in order of kolmogorov complexity, the simplest most elegant solution comes up first, by default.
<0> and doing a simple search does that easily
<0> dmiles_afk, slowly but surely
<0> dmiles_afk, improving my tools and environment
<5> lots of right clicks?
<5> erm context menu feature
<5> s
<0> not so much
<0> i'm trying to keep the keyboard shortcuts as much as possible :-)
<0> ttt-, you don't need to understand specific program instances if you understand how to create them instead.
<0> ttt-, externally, refactoring behaves just like a complete rewrite :-)
<3> <daYZman@ef> hi
<3> <daYZman@ef> does anyone know why research on TDT (topic detection and tracking) seems to be so rare?
<0> daYZman, what context? text mining?
<3> <daYZman@ef> yes
<0> where did you search?
<3> <daYZman@ef> it seems to me most universities don't seem to have an interest in TDT.
<0> have you searched citeseer?
<0> most universities are not interested in a majority of specialised research topics... only few unis specialize
<0> well, many unis specialize, but each in different subjects :-)
<3> <daYZman@ef> yeah i searched citeseer. with topic detection, it returns like 300 results
<2> topic detection?
<0> text mining
<3> <daYZman@ef> well.. if i search topic detection tracking, i get 20
<0> daYZman, type in this channel on freenode, it's easier :-)
<6> hi
<0> hi ;)
<7> moin
<0> daYZman, i presume you need to find a different and related topic name
<0> hey bourbaki
<6> the reason i ask is that i'm kinda interested in this idea of TDT, and thinking of pursuing a phd on it
<6> but then, it looks so rare
<7> tdt?
<2> why are you interested in text AI
<6> bourbaki, topic detection and tracking
<2> it's really hard :/
<0> bourbaki, text mining
<2> like the topic of a text?
<0> daYZman, well, if it wasn't rare, you'd have more trouble getting a Ph.D. :-)
<7> ah
<2> daYZman: do you have new ideas to do this?
<0> columbia has a lot of NLP/NLG
<7> AJC did you check out mindrover?
<0> newsblaster or something
<0> bourbaki, briefly, didn't download it
<6> ttt-, well, i have some ideas..
<2> like what?
<0> don't say the word search :-)
<0> ttt- freaks out!
<2> or statistical methods
Return to
#ai
or
Go to some related
logs:
#oe
comparing string vars javascript trim
#web
#linux
HOWTO install apache2 on ubuntu
linux hybrid sleep -vista
#centos
Broken runlevel entry
chkrootkit PACKET SNIFFER eth
fluxspace undefined symbol
|
|