Loading ...
Sorry, an error occurred while loading the content.

RE: [GP] analog ckt design question / Depth-first appears to be more crossover friendly

Expand Messages
  • jon
    now this is cool! the next best thing that could happen to me would be a free download of your books! =P ... actually i found breadth-first to be appropriate
    Message 1 of 4 , Apr 14, 2003
    • 0 Attachment
      now this is cool! the next best thing that could
      happen to me would be a free download of your books!
      =P

      --- John Koza <john@...> wrote:

      > When we first started doing automatic synthesis of
      > electrical circuits by
      > means of genetic programming in 1995, we used a
      > breath-first order of
      > evaluation during the developmental process.
      > However, we have come to
      > believe that this choice is not as efficient as the
      > obvious alternative
      > (i.e., depth-first evaluation).
      actually i found breadth-first to be appropriate only
      when there is that function where a write-head is
      introduced to two nodes that are at least 2 braches
      away (i remember there is such in your ieee paper)

      > The genetic algorithm and genetic programming work
      > on the principle that
      > there is something about the structure of a
      > relatively fit individual that
      > contributes to the individual's fitness. In genetic
      > programming, the
      > crossover operation creates new individuals by
      > recombining subtrees of
      > relatively fit individuals. Thus, it seems
      > reasonable that the crossover
      > operation would be more efficient if the parts of a
      > program that are moved
      > by the crossover operation (i.e., subtrees) more
      > closely corresponded to the
      > sequence of component-creating and
      > topology-modifying functions.
      i'm lost here. isn't it that the evaluation is done
      only to determine its fitness, independent of
      population mumbo-jumbo activities?


      > That is, we think it is desirable to preserve the
      > locality inherent in the
      > subtrees.
      hmmm, so the crossover operator should be tailor-made
      a little because not all points in the tree can be a
      crossover point?


      thanks
      jon


      __________________________________________________
      Do you Yahoo!?
      The New Yahoo! Search - Faster. Easier. Bingo
      http://search.yahoo.com
    • jon
      hello, might as well mail the list, i find it quite quiet. sorry, i m quite confused on how replies to a list works here... ... i ve just started for a month
      Message 2 of 4 , Apr 15, 2003
      • 0 Attachment
        hello,

        might as well mail the list, i find it quite quiet.
        sorry, i'm quite confused on how replies to a list
        works here...

        --- Varun Aggarwal <varun.aggarwal@...> wrote:
        > Hi!
        >
        > Good to hear from you. Its nice to see someone
        > working on what Koza did,
        > infact, I didnt read any significant work carried
        > out on the approach used
        > by Koza for analog synthesis after he did the
        > initial work, it has been
        > dormant for a while!!!
        i've just started for a month on this, a

        > Well, The maximum elements will be fixed in linear
        > encoding (however C. Goh
        > even tackles this in his paper by starting from
        the
        > minimum size and
        do you have those gecco papers? i also want to read
        grimbleby's stuff

        > component is being used. However, I believe, same
        > will be true with trees as
        > well, where you will have a maximum depth or
        maximum
        > number of nodes as a
        > parameter.
        hmmm, you have a good point.

        actually i was trying to modify lilgp here and i
        read
        the docs (btw, relatively great docs about
        implementation), and i was impressed. the lilgp
        author
        said that trees were represented as arrays with
        structuring "signals" embedded, against the more
        intuitive way of using pointers to connect nodes in
        the tree. this is good, he said, because after many
        generations of crossover an individual's node may be
        scattered around the process' memory space. now, if
        the app actually uses a lot of memory, the os may
        swap
        memory to disk and this scattered state would cause
        it
        to thrash. it's low level, makes sense, but i'd like
        to see hard graphs!

        > Although, you should also look at the
        representation
        > used by Lohn,
        > Colambano.
        i had one of them from citeseer. actually i'm quite
        worried because in that paper the results were quite
        simple configurations while in koza's were very
        intimidating for a lowpass filter

        > Yeh! Pipelining is the easiest approach. Its very
        > easy to do it using PSPICE
        windows have pipes? i heard about dde in the past
        but
        i never figured. cool!

        > Best of luck for your project!!! Once you are near
        thanks mate, i need tons of it! =P

        > interested, may be we can work jointly on trying
        my
        > Learning Scheme (see my
        > homepage) on your algorithm to see if there are
        i proposed on lisp because it has a larger library,
        although i find scheme more neat.

        > better results. I have
        > already used it with linear representation on the
        > oscillator synthesis
        > problem, where it gave better results.
        better results against?

        thanks
        jon


        __________________________________________________
        Do you Yahoo!?
        The New Yahoo! Search - Faster. Easier. Bingo
        http://search.yahoo.com
      Your message has been successfully submitted and would be delivered to recipients shortly.