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

Re: [GP] Self Organization and scalable solutions with emergent phenomena

Expand Messages
  • dennis petkiewicz
    Ian... Wanted to thank you for your thoughtful and very insightful reply. I will look for the post you referenced and the additional research material. It
    Message 1 of 12 , Dec 31, 2003
    • 0 Attachment
      Ian...

      Wanted to thank you for your thoughtful and very
      insightful reply. I will look for the post you
      referenced and the additional research material.

      It sounds that the framework for the "complete model"
      could be organized as the interaction of three
      domains:
      1) the elements (DNA material..molecules, formulas,
      terminals...stuff) that gets manipulated
      2) the algorithm (or the laws, properties, behavior)
      that manages the combination, interaction and
      synthesis of the elements.
      3) the environment (or the physical or logical place
      where the elements and algorithm interact)

      Seems we have a good handle on #1 but the algorithm
      and enabling environment is where the discovery is.

      In looking at GA GP and GEP techniques it seems that
      we are making an assumption that the algorithm is
      hierarchical. From your discussion it looks as if
      there may be a 2nd or 3rd order structure...the
      interaction of the phonemes (and proteins, RNA, etc)
      to evolve; and the ability of the phonemes to in-turn
      modify the genetic code.

      Are you aware of any genetic algorithms that utilize
      these properties?

      --- Ian Badcoe <ian_badcoe@...> wrote:
      > At 14:09 09/12/2003 -0800, you wrote:
      > >Ivan, your observation that...
      > >"Perhaps the
      > >self-organization of genotypic instructions into
      > >phenotypes is a key
      > >missing ingredient necessary for unleashing the
      > >evolution of complex
      > >and scalable solutions with emergent phenomena such
      > >as: scale-free-ness,
      > >adaptability, innovation, evolvability, and
      > >robustness. This workshop
      > >will focus on domain-independent methods for
      > >representing complex
      > >solutions with relatively simple self-organizable
      > >building blocks."
      > >
      > >Is very insightful...and a concept I had referred
      > to
      > >earlier as the "the complete model" Where can I
      > >obtain research material progressive in this area?
      >
      > I don't know, but if you get answers please share
      > them.
      >
      > >As I had replied earlier under the topic "Why
      > >Genetics?":
      > >
      > >Seems that if we had one, and could emulate it, we
      > >would have a very powerful thing: just look at what
      > >mother nature did. How complex organisms came to
      > >exist without an explicit pre-existing design, and
      > no
      > >awareness of physical laws and the nature of matter
      > &
      > >energy is astounding. The ability to replicate this
      > >ability on demand would be revolutionary.
      >
      > Yes and no. I used to be biochemist/genetic
      > engineer, so I feel I
      > have a pretty good field for both areas here (GP and
      > "real genetics").
      >
      > Nature has produced a pretty potent evolution
      > system, but a few
      > warnings come into play when you compare it with GP:
      >
      > (I don't like using the word "nature" in this way,
      > as if it were a thing
      > or a system, or anything more than a broad
      > collection of only
      > partly-related concepts, but for the sake of
      > simple phrasing...)
      >
      > 1) Nature operates over very large scales compared
      > to GP, anybody ever
      > used a population size of a trillion?
      > Viruses can do that in 1
      > litre of
      > water. Bacteria can manage a population of
      > a hundred billion in the
      > human gut.
      >
      > 2) Nature's genomes tend to be large than GP's, the
      > human genome is
      > about 3/4 of a gigabyte, anybody every used
      > a genome that big?
      >
      > 3) Nature has longer to play with, rats and mice
      > became different species
      > (depending how you estimate it) about 10
      > million years ago. That's
      > between 10 and 20 million generations, just
      > to accumulate the
      > differences between rats and mice
      >
      > Now, these are all differences of quantity, not
      > quality, but they do pretty
      > clearly show that nature is playing in a different
      > ball-park from us humble
      > software engineers... and they all raise the
      > warning that any technique
      > found in nature may be valid but impractical for our
      > purposes.
      >
      > --
      >
      > OK so that's the warnings, it's also useful to take
      > some account
      > all that other stuff which Gordon Pusch brought up,
      > but I might make a
      > separate mail on that. However, after all the
      > warnings have been taken
      > into account, there are some useful hints and
      > suggestions to be
      > extracted from natural genetics:
      >
      > << as a reminder to non-biochemists:
      >
      > DNA encodes RNA encodes Protein.
      > Proteins are a 1D sequence of monomers, but
      > in water spontaneously
      > fold into precise 3D structures.
      > Some proteins catalyze chemical reactions
      > and are called enzymes.
      > Proteins can bind (stick to) other
      > proteins, DNA or small molcules,
      > either with great selectivity or target or more
      > promiscuously.
      >
      > >>
      >
      > A) Parallel processing
      >
      > Natural genetics has two very powerful
      > mechanisms of parallelism
      >
      > A-1) Multiple genes
      >
      > The separate genes in an organism act to
      > divide each problem into
      > smaller pieces (although for nature, note that the
      > existence of these genes
      > is part of the solution to the problem not a
      > pre-condition of the
      > search). Thus,
      > the problem "how do we metabolize sugars" is
      > addressed by a group of about
      > 30 genes (= 30 enzymes) each having only to handle a
      > single, simple reaction
      > (and control of the 30 enzymes is layered on top in
      > a manner I'll get to in
      > a bit)
      >
      > A-2) Each gene product acts in parallel.
      >
      > E.g. it is a simple as the enzyme coded by
      > each gene is dropped into
      > the same bath of solvent and they all compete with
      > one another for the
      > substrates they act on. Thus the addition of a new
      > enzyme requires no
      > special "wiring" (or in software terms --
      > initialization) to connect it to the
      > rest of the system. It's just dropped into the pot
      > with all the rest.
      >
      > Some notable exceptions to this exist, e.g.
      > the partition of the
      > cytoplasm from the mitochondrion interior, but these
      > are take further
      > advantage of the parallelism (e.g. by running two
      > parallel environments
      > with separate ground-rules)
      >
      > B) Weak coupling
      >
      > There are many examples of weakly coupled
      > control mechanisms
      > in natural organisms. The meaning of "weakly
      > coupled" is that it is relatively
      > easy to rebind the connection between the controlled
      > and the controller.
      > The importance of weak coupling is that it enables
      > existing, fragile, and
      > hard to evolve systems to be re-targeted without
      > disrupting their insides.
      > An analogy might be to a well-designed API, with a
      > very complex and
      > finely-tuned interior, but which can be trivially
      > plugged into any application,
      > because the details of the using code are insulated
      > from it.
      >
      > B-1) Genes are controlled by adjoining control
      > sequences on the DNA. The
      > gene can be readily moved to adjoin different
      > sequences, or control sequences
      > can mutate without damaging the gene.
      >
      > B-2) Control-sequences are acted on by the
      > products of other genes, which
      > can mutate or have their own controls changed
      >
      > B-3) Enzymes can be regulated by binding small
      > molecules or other proteins
      > at sites remote from their "business end". Thus the
      > control of a single enzyme
      > can evolve without affecting it's function. If the
      > controlling molecule is
      > another
      > protein then there is even more flexibility as the
      > controlling protein can
      > evolve
      > in ways which affect its binding to the target, or
      > to be itself controlled by
      > something binding to it.
      >
      > B-4) Hormones bind to one end of a receptor,
      > typically triggering an
      > effect at
      > the opposite end...
      >
      === message truncated ===


      __________________________________
      Do you Yahoo!?
      Free Pop-Up Blocker - Get it now
      http://companion.yahoo.com/
    • David vun Kannon
      Hi, I think you mean phenotype when you wrote phoneme. Phenotype is just a fancy word for the general body plan that can be constructed from the genotype - the
      Message 2 of 12 , Jan 4, 2004
      • 0 Attachment
        Hi,
         
        I think you mean phenotype when you wrote phoneme. Phenotype is just a fancy word for the general body plan that can be constructed from the genotype - the genetic instructions. I sometimes find it difficult to identify the "phenotype" in GA/GP if the structure manipulated by the evolutionary algorithm is the genotype. It seems that many times the genotype is interacting directly with the environment.
         
        There are many applications of GP in which individuals build phenotypes that then interact with each other as well as the environment. Co-evolution is an obvious example. GP to evolve NN weights that drive robot behaviors in multi-agent settings is another.
         
        The only way for phenotypes to influence the pool of available genotypes is to survive and breed. They cannot modify the genetic code directly - that is the error of Lamarck, the idea that giraffes get longer necks by trying to reach higher leaves. It is not part of the "complete model" of "variation within a population and differential breeding of individuals according to their success within the environment".
         
        IMHO, one of the difficulties of credit assignment algorithms such as back propagation in neural nets and bucket brigade in classifier systems is that they attempt a Lamarckian learning which is inherently problematic.
        ----- Original Message -----
        Sent: Wednesday, December 31, 2003 12:23 PM
        Subject: Re: [GP] Self Organization and scalable solutions with emergent phenomena

        Ian...

        Wanted to thank you for your thoughtful and very
        insightful reply. I will look for the post you
        referenced and the additional research material.

        It sounds that the framework for the  "complete model"
        could be organized as the interaction of three
        domains:
        1) the elements (DNA material..molecules, formulas,
        terminals...stuff) that gets manipulated
        2) the algorithm (or the laws, properties, behavior)
        that manages the combination, interaction and
        synthesis of the elements.
        3) the environment (or the physical or logical place
        where the elements and algorithm interact)

        Seems we have a good handle on #1 but the algorithm
        and enabling environment is where the discovery is.

        In looking at GA GP and GEP techniques it seems that
        we are making an assumption that the algorithm is
        hierarchical.  From your discussion it looks as if
        there may be a 2nd or 3rd order structure...the
        interaction of the phonemes (and proteins, RNA, etc)
        to evolve; and the ability of the phonemes to in-turn
        modify the genetic code. 

        Are you aware of any genetic algorithms that utilize
        these properties?

        --- Ian Badcoe <ian_badcoe@...> wrote:
        > At 14:09 09/12/2003 -0800, you wrote:
        > >Ivan, your observation that...
        > >"Perhaps the
        > >self-organization of genotypic instructions into
        > >phenotypes is a key
        > >missing ingredient necessary for unleashing the
        > >evolution of complex
        > >and scalable solutions with emergent phenomena such
        > >as: scale-free-ness,
        > >adaptability, innovation, evolvability, and
        > >robustness. This workshop
        > >will focus on domain-independent methods for
        > >representing complex
        > >solutions with relatively simple self-organizable
        > >building blocks."
        > >
        > >Is very insightful...and a concept I had referred
        > to
        > >earlier as the "the complete model"  Where can I
        > >obtain research material progressive in this area?
        >
        > I don't know, but if you get answers please share
        > them.
        >
        > >As I had replied earlier under the topic "Why
        > >Genetics?":
        > >
        > >Seems that if we had one, and could emulate it, we
        > >would have a very powerful thing: just look at what
        > >mother nature did.  How complex organisms came to
        > >exist without an explicit pre-existing design, and
        > no
        > >awareness of physical laws and the nature of matter
        > &
        > >energy is astounding. The ability to replicate this
        > >ability on demand would be revolutionary.
        >
        > Yes and no.  I used to be biochemist/genetic
        > engineer, so I feel I
        > have a pretty good field for both areas here (GP and
        > "real genetics").
        >
        > Nature has produced a pretty potent evolution
        > system, but a few
        > warnings come into play when you compare it with GP:
        >
        > (I don't like using the word "nature" in this way,
        > as if it were a thing
        >   or a system, or anything more than a broad
        > collection of only
        >   partly-related concepts, but for the sake of
        > simple phrasing...)
        >
        > 1) Nature operates over very large scales compared
        > to GP, anybody ever
        >          used a population size of a trillion?
        > Viruses can do that in 1
        > litre of
        >          water.  Bacteria can manage a population of
        > a hundred billion in the
        >          human gut.
        >
        > 2) Nature's genomes tend to be large than GP's, the
        > human genome is
        >          about 3/4 of a gigabyte, anybody every used
        > a genome that big?
        >
        > 3) Nature has longer to play with, rats and mice
        > became different species
        >          (depending how you estimate it) about 10
        > million years ago.  That's
        >          between 10 and 20 million generations, just
        > to accumulate the
        >          differences between rats and mice
        >
        > Now, these are all differences of quantity, not
        > quality, but they do pretty
        > clearly show that nature is playing in a different
        > ball-park from us humble
        > software engineers...  and they all raise the
        > warning that any technique
        > found in nature may be valid but impractical for our
        > purposes.
        >
        > --
        >
        > OK so that's the warnings, it's also useful to take
        > some account
        > all that other stuff which Gordon Pusch brought up,
        > but I might make a
        > separate mail on that.  However, after all the
        > warnings have been taken
        > into account, there are some useful hints and
        > suggestions to be
        > extracted from natural genetics:
        >
        > << as a reminder to non-biochemists:
        >
        >          DNA encodes RNA encodes Protein.
        >          Proteins are a 1D sequence of monomers, but
        > in water spontaneously
        > fold into precise 3D structures.
        >          Some proteins catalyze chemical reactions
        > and are called enzymes.
        >          Proteins can bind (stick to) other
        > proteins, DNA or small molcules,
        > either with great selectivity or target or more
        > promiscuously.
        >
        >  >>
        >
        > A)      Parallel processing
        >
        >          Natural genetics has two very powerful
        > mechanisms of parallelism
        >
        > A-1)    Multiple genes
        >
        >          The separate genes in an organism act to
        > divide each problem into
        > smaller pieces (although for nature, note that the
        > existence of these genes
        > is part of the solution to the problem not a
        > pre-condition of the
        > search).  Thus,
        > the problem "how do we metabolize sugars" is
        > addressed by a group of about
        > 30 genes (= 30 enzymes) each having only to handle a
        > single, simple reaction
        > (and control of the 30 enzymes is layered on top in
        > a manner I'll get to in
        > a bit)
        >
        > A-2)    Each gene product acts in parallel.
        >
        >          E.g. it is a simple as the enzyme coded by
        > each gene is dropped into
        > the same bath of solvent and they all compete with
        > one another for the
        > substrates they act on.  Thus the addition of a new
        > enzyme requires no
        > special "wiring" (or in software terms --
        > initialization) to connect it to the
        > rest of the system.  It's just dropped into the pot
        > with all the rest.
        >
        >          Some notable exceptions to this exist, e.g.
        > the partition of the
        > cytoplasm from the mitochondrion interior, but these
        > are take further
        > advantage of the parallelism (e.g. by running two
        > parallel environments
        > with separate ground-rules)
        >
        > B)      Weak coupling
        >
        >          There are many examples of weakly coupled
        > control mechanisms
        > in natural organisms.  The meaning of "weakly
        > coupled" is that it is relatively
        > easy to rebind the connection between the controlled
        > and the controller.
        > The importance of weak coupling is that it enables
        > existing, fragile, and
        > hard to evolve systems to be re-targeted without
        > disrupting their insides.
        > An analogy might be to a well-designed API, with a
        > very complex and
        > finely-tuned interior, but which can be trivially
        > plugged into any application,
        > because the details of the using code are insulated
        > from it.
        >
        > B-1)    Genes are controlled by adjoining control
        > sequences on the DNA.  The
        > gene can be readily moved to adjoin different
        > sequences, or control sequences
        > can mutate without damaging the gene.
        >
        > B-2)    Control-sequences are acted on by the
        > products of other genes, which
        > can mutate or have their own controls changed
        >
        > B-3)    Enzymes can be regulated by binding small
        > molecules or other proteins
        > at sites remote from their "business end".  Thus the
        > control of a single enzyme
        > can evolve without affecting it's function.  If the
        > controlling molecule is
        > another
        > protein then there is even more flexibility as the
        > controlling protein can
        > evolve
        > in ways which affect its binding to the target, or
        > to be itself controlled by
        > something binding to it.
        >
        > B-4)    Hormones bind to one end of a receptor,
        > typically triggering an
        > effect at
        > the opposite end...
        >
        === message truncated ===


        __________________________________
        Do you Yahoo!?
        Free Pop-Up Blocker - Get it now
        http://companion.yahoo.com/


        Yahoo! Groups Links

      • dennis petkiewicz
        David...thanks for the clarification and insight. It still seems, though that the ability to evolve is a desirable trait; and an organism that can evolve and
        Message 3 of 12 , Jan 4, 2004
        • 0 Attachment
          David...thanks for the clarification and insight.

          It still seems, though that the ability to evolve is a
          desirable trait; and an organism that can evolve and
          therefore adapt to change would have an advantage, and
          therefore reproduce. Think the HIV virus would be an
          example of this.

          Regards,

          Dennis Petkiewicz


          --- David vun Kannon <dvunkannon@...> wrote:
          > Hi,
          >
          > I think you mean phenotype when you wrote phoneme.
          > Phenotype is just a fancy word for the general body
          > plan that can be constructed from the genotype - the
          > genetic instructions. I sometimes find it difficult
          > to identify the "phenotype" in GA/GP if the
          > structure manipulated by the evolutionary algorithm
          > is the genotype. It seems that many times the
          > genotype is interacting directly with the
          > environment.
          >
          > There are many applications of GP in which
          > individuals build phenotypes that then interact with
          > each other as well as the environment. Co-evolution
          > is an obvious example. GP to evolve NN weights that
          > drive robot behaviors in multi-agent settings is
          > another.
          >
          > The only way for phenotypes to influence the pool of
          > available genotypes is to survive and breed. They
          > cannot modify the genetic code directly - that is
          > the error of Lamarck, the idea that giraffes get
          > longer necks by trying to reach higher leaves. It is
          > not part of the "complete model" of "variation
          > within a population and differential breeding of
          > individuals according to their success within the
          > environment".
          >
          > IMHO, one of the difficulties of credit assignment
          > algorithms such as back propagation in neural nets
          > and bucket brigade in classifier systems is that
          > they attempt a Lamarckian learning which is
          > inherently problematic.
          > ----- Original Message -----
          > From: dennis petkiewicz
          > To: genetic_programming@yahoogroups.com ;
          > ian_badcoe@...
          > Sent: Wednesday, December 31, 2003 12:23 PM
          > Subject: Re: [GP] Self Organization and scalable
          > solutions with emergent phenomena
          >
          >
          > Ian...
          >
          > Wanted to thank you for your thoughtful and very
          > insightful reply. I will look for the post you
          > referenced and the additional research material.
          >
          > It sounds that the framework for the "complete
          > model"
          > could be organized as the interaction of three
          > domains:
          > 1) the elements (DNA material..molecules,
          > formulas,
          > terminals...stuff) that gets manipulated
          > 2) the algorithm (or the laws, properties,
          > behavior)
          > that manages the combination, interaction and
          > synthesis of the elements.
          > 3) the environment (or the physical or logical
          > place
          > where the elements and algorithm interact)
          >
          > Seems we have a good handle on #1 but the
          > algorithm
          > and enabling environment is where the discovery
          > is.
          >
          > In looking at GA GP and GEP techniques it seems
          > that
          > we are making an assumption that the algorithm is
          > hierarchical. From your discussion it looks as if
          > there may be a 2nd or 3rd order structure...the
          > interaction of the phonemes (and proteins, RNA,
          > etc)
          > to evolve; and the ability of the phonemes to
          > in-turn
          > modify the genetic code.
          >
          > Are you aware of any genetic algorithms that
          > utilize
          > these properties?
          >
          > --- Ian Badcoe <ian_badcoe@...> wrote:
          > > At 14:09 09/12/2003 -0800, you wrote:
          > > >Ivan, your observation that...
          > > >"Perhaps the
          > > >self-organization of genotypic instructions
          > into
          > > >phenotypes is a key
          > > >missing ingredient necessary for unleashing the
          > > >evolution of complex
          > > >and scalable solutions with emergent phenomena
          > such
          > > >as: scale-free-ness,
          > > >adaptability, innovation, evolvability, and
          > > >robustness. This workshop
          > > >will focus on domain-independent methods for
          > > >representing complex
          > > >solutions with relatively simple
          > self-organizable
          > > >building blocks."
          > > >
          > > >Is very insightful...and a concept I had
          > referred
          > > to
          > > >earlier as the "the complete model" Where can
          > I
          > > >obtain research material progressive in this
          > area?
          > >
          > > I don't know, but if you get answers please
          > share
          > > them.
          > >
          > > >As I had replied earlier under the topic "Why
          > > >Genetics?":
          > > >
          > > >Seems that if we had one, and could emulate it,
          > we
          > > >would have a very powerful thing: just look at
          > what
          > > >mother nature did. How complex organisms came
          > to
          > > >exist without an explicit pre-existing design,
          > and
          > > no
          > > >awareness of physical laws and the nature of
          > matter
          > > &
          > > >energy is astounding. The ability to replicate
          > this
          > > >ability on demand would be revolutionary.
          > >
          > > Yes and no. I used to be biochemist/genetic
          > > engineer, so I feel I
          > > have a pretty good field for both areas here (GP
          > and
          > > "real genetics").
          > >
          > > Nature has produced a pretty potent evolution
          > > system, but a few
          > > warnings come into play when you compare it with
          > GP:
          > >
          > > (I don't like using the word "nature" in this
          > way,
          > > as if it were a thing
          > > or a system, or anything more than a broad
          > > collection of only
          > > partly-related concepts, but for the sake of
          > > simple phrasing...)
          > >
          > > 1) Nature operates over very large scales
          > compared
          > > to GP, anybody ever
          > > used a population size of a trillion?
          > > Viruses can do that in 1
          > > litre of
          > > water. Bacteria can manage a
          > population of
          > > a hundred billion in the
          > > human gut.
          > >
          > > 2) Nature's genomes tend to be large than GP's,
          > the
          > > human genome is
          > > about 3/4 of a gigabyte, anybody every
          > used
          > > a genome that big?
          > >
          > > 3) Nature has longer to play with, rats and mice
          > > became different species
          > > (depending how you estimate it) about
          > 10
          > > million years ago. That's
          > > between 10 and 20 million generations,
          > just
          > > to accumulate the
          > > differences between rats and mice
          > >
          > > Now, these are all differences of quantity, not
          > > quality, but they do pretty
          > > clearly show that nature is playing in a
          > different
          > > ball-park from us humble
          > > software engineers... and they all raise the
          > > warning that any technique
          > > found in nature may be valid but impractical for
          > our
          > > purposes.
          > >
          > > --
          > >
          > > OK so that's the warnings, it's also useful to
          > take
          > > some account
          > > all that other stuff which Gordon Pusch brought
          > up,
          > > but I might make a
          >
          === message truncated ===


          __________________________________
          Do you Yahoo!?
          Free Pop-Up Blocker - Get it now
          http://companion.yahoo.com/
        • David vun Kannon
          At least in nature, the ability to evolve is a trait of populations, while the ability to learn is a trait of individuals. When the two are combined (as in
          Message 4 of 12 , Jan 4, 2004
          • 0 Attachment
            At least in nature, the ability to evolve is a trait of populations, while the ability to learn is a trait of individuals. When the two are combined (as in humanity) they are quite powerful. Your example of a retrovirus is interesting because (as I said about some GP applications) there doesn't seem to be a body interposed between the genotype and the environment. I don't know enough about HIV et al. to say definitely, but I doubt that individual viral particles are changed by their activities within their environment and pass those changes on when they hack the machinery of a cell and force it to reproduce more virii. I don't know how virii mutate.
            ----- Original Message -----
            Sent: Sunday, January 04, 2004 6:45 PM
            Subject: Re: [GP] Self Organization and scalable solutions with emergent phenomena

            David...thanks for the clarification and insight.

            It still seems, though that the ability to evolve is a
            desirable trait; and an organism that can evolve and
            therefore adapt to change would have an advantage, and
            therefore reproduce.  Think the HIV virus would be an
            example of this. 

            Regards,

            Dennis Petkiewicz


            --- David vun Kannon <dvunkannon@...> wrote:
            > Hi,
            >
            > I think you mean phenotype when you wrote phoneme.
            > Phenotype is just a fancy word for the general body
            > plan that can be constructed from the genotype - the
            > genetic instructions. I sometimes find it difficult
            > to identify the "phenotype" in GA/GP if the
            > structure manipulated by the evolutionary algorithm
            > is the genotype. It seems that many times the
            > genotype is interacting directly with the
            > environment.
            >
            > There are many applications of GP in which
            > individuals build phenotypes that then interact with
            > each other as well as the environment. Co-evolution
            > is an obvious example. GP to evolve NN weights that
            > drive robot behaviors in multi-agent settings is
            > another.
            >
            > The only way for phenotypes to influence the pool of
            > available genotypes is to survive and breed. They
            > cannot modify the genetic code directly - that is
            > the error of Lamarck, the idea that giraffes get
            > longer necks by trying to reach higher leaves. It is
            > not part of the "complete model" of "variation
            > within a population and differential breeding of
            > individuals according to their success within the
            > environment".
            >
            > IMHO, one of the difficulties of credit assignment
            > algorithms such as back propagation in neural nets
            > and bucket brigade in classifier systems is that
            > they attempt a Lamarckian learning which is
            > inherently problematic.
            >   ----- Original Message -----
            >   From: dennis petkiewicz
            >   To: genetic_programming@yahoogroups.com ;
            > ian_badcoe@...
            >   Sent: Wednesday, December 31, 2003 12:23 PM
            >   Subject: Re: [GP] Self Organization and scalable
            > solutions with emergent phenomena
            >
            >
            >   Ian...
            >
            >   Wanted to thank you for your thoughtful and very
            >   insightful reply. I will look for the post you
            >   referenced and the additional research material.
            >
            >   It sounds that the framework for the  "complete
            > model"
            >   could be organized as the interaction of three
            >   domains:
            >   1) the elements (DNA material..molecules,
            > formulas,
            >   terminals...stuff) that gets manipulated
            >   2) the algorithm (or the laws, properties,
            > behavior)
            >   that manages the combination, interaction and
            >   synthesis of the elements.
            >   3) the environment (or the physical or logical
            > place
            >   where the elements and algorithm interact)
            >
            >   Seems we have a good handle on #1 but the
            > algorithm
            >   and enabling environment is where the discovery
            > is.
            >
            >   In looking at GA GP and GEP techniques it seems
            > that
            >   we are making an assumption that the algorithm is
            >   hierarchical.  From your discussion it looks as if
            >   there may be a 2nd or 3rd order structure...the
            >   interaction of the phonemes (and proteins, RNA,
            > etc)
            >   to evolve; and the ability of the phonemes to
            > in-turn
            >   modify the genetic code. 
            >
            >   Are you aware of any genetic algorithms that
            > utilize
            >   these properties?
            >
            >   --- Ian Badcoe <ian_badcoe@...> wrote:
            >   > At 14:09 09/12/2003 -0800, you wrote:
            >   > >Ivan, your observation that...
            >   > >"Perhaps the
            >   > >self-organization of genotypic instructions
            > into
            >   > >phenotypes is a key
            >   > >missing ingredient necessary for unleashing the
            >   > >evolution of complex
            >   > >and scalable solutions with emergent phenomena
            > such
            >   > >as: scale-free-ness,
            >   > >adaptability, innovation, evolvability, and
            >   > >robustness. This workshop
            >   > >will focus on domain-independent methods for
            >   > >representing complex
            >   > >solutions with relatively simple
            > self-organizable
            >   > >building blocks."
            >   > >
            >   > >Is very insightful...and a concept I had
            > referred
            >   > to
            >   > >earlier as the "the complete model"  Where can
            > I
            >   > >obtain research material progressive in this
            > area?
            >   >
            >   > I don't know, but if you get answers please
            > share
            >   > them.
            >   >
            >   > >As I had replied earlier under the topic "Why
            >   > >Genetics?":
            >   > >
            >   > >Seems that if we had one, and could emulate it,
            > we
            >   > >would have a very powerful thing: just look at
            > what
            >   > >mother nature did.  How complex organisms came
            > to
            >   > >exist without an explicit pre-existing design,
            > and
            >   > no
            >   > >awareness of physical laws and the nature of
            > matter
            >   > &
            >   > >energy is astounding. The ability to replicate
            > this
            >   > >ability on demand would be revolutionary.
            >   >
            >   > Yes and no.  I used to be biochemist/genetic
            >   > engineer, so I feel I
            >   > have a pretty good field for both areas here (GP
            > and
            >   > "real genetics").
            >   >
            >   > Nature has produced a pretty potent evolution
            >   > system, but a few
            >   > warnings come into play when you compare it with
            > GP:
            >   >
            >   > (I don't like using the word "nature" in this
            > way,
            >   > as if it were a thing
            >   >   or a system, or anything more than a broad
            >   > collection of only
            >   >   partly-related concepts, but for the sake of
            >   > simple phrasing...)
            >   >
            >   > 1) Nature operates over very large scales
            > compared
            >   > to GP, anybody ever
            >   >          used a population size of a trillion?
            >   > Viruses can do that in 1
            >   > litre of
            >   >          water.  Bacteria can manage a
            > population of
            >   > a hundred billion in the
            >   >          human gut.
            >   >
            >   > 2) Nature's genomes tend to be large than GP's,
            > the
            >   > human genome is
            >   >          about 3/4 of a gigabyte, anybody every
            > used
            >   > a genome that big?
            >   >
            >   > 3) Nature has longer to play with, rats and mice
            >   > became different species
            >   >          (depending how you estimate it) about
            > 10
            >   > million years ago.  That's
            >   >          between 10 and 20 million generations,
            > just
            >   > to accumulate the
            >   >          differences between rats and mice
            >   >
            >   > Now, these are all differences of quantity, not
            >   > quality, but they do pretty
            >   > clearly show that nature is playing in a
            > different
            >   > ball-park from us humble
            >   > software engineers...  and they all raise the
            >   > warning that any technique
            >   > found in nature may be valid but impractical for
            > our
            >   > purposes.
            >   >
            >   > --
            >   >
            >   > OK so that's the warnings, it's also useful to
            > take
            >   > some account
            >   > all that other stuff which Gordon Pusch brought
            > up,
            >   > but I might make a
            >
            === message truncated ===


            __________________________________
            Do you Yahoo!?
            Free Pop-Up Blocker - Get it now
            http://companion.yahoo.com/


            Yahoo! Groups Links

          • dennis petkiewicz
            David... please consider elaborating on your thoughts on why one of the difficulties of credit assignment ... Would like to have this background as a
            Message 5 of 12 , Jan 6, 2004
            • 0 Attachment
              David... please consider elaborating on your thoughts
              on why "one of the difficulties of credit assignment
              > algorithms such as back propagation in neural nets
              > and bucket brigade in classifier systems is that
              > they attempt a Lamarckian learning which is
              > inherently problematic."

              Would like to have this background as a 'heads-up' in
              looking into the workings of these other processes.

              Most finest regards,
              dennis



              __________________________________
              Do you Yahoo!?
              Yahoo! Hotjobs: Enter the "Signing Bonus" Sweepstakes
              http://hotjobs.sweepstakes.yahoo.com/signingbonus
            • David vun Kannon
              Dennis, In a stimulus-response system (no hidden layers of neurons) error correction (credit allocation with a switched sign) is easy. There are N direct
              Message 6 of 12 , Jan 6, 2004
              • 0 Attachment
                Dennis,
                 
                In a stimulus-response system (no hidden layers of neurons) error correction (credit allocation with a switched sign) is easy. There are N direct inputs to neuron, and if it fired when it should have been quiet, those input weights can be seen to be directly responsible and therefore adjusted. But, since S-R perceptrons can't calculate an XOR, much less anything more interesting, we need to add hidden layers of neurons to achieve interesting computational power. But the downside is a combinatorial explosion in the number of possible weights and activation strengths that can lead to a neuron in the output layer firing.
                 
                Similarly, if an individual survives to breed, to which part of the genome should credit be differentially given? How should the long necked giraffe, having eaten low and high leaves all its life (but more high ones than other giraffes!) discover that its differential survival and breeding capacity was caused by its ability to reach higher leaves? How can that awareness, if it could exist, be translated into a knowledge of all the parts of the genome which influence all the developmental pathways that result in the longer limbs and bigger heart and stronger blood vessels (to pump blood up so long a neck) which give it a high leaf eating capacity in excess of its neighbors. This is like one neuron (breed/die) firing after an uncountable number of hidden layers of other neurons separating it from the input layer (the genome), layers of neurons representing (and I know I'm straining the analogy) the activation of developmental paths in the creation of the body and the entire history of the body until breeding or death. And not just its own genome and body, but those of its neighbors, because survival is not doing well, it is just doing better.
                As impossible as that is when only one trait is varying, it is all the more impossible when thousands of traits and behaviors are varying all at the same time.
                 
                So Lamarckianism is an extreme example of fighting against a combinatorial explosion in the number of prior events that could be responsible for a particular event. Backprop and bucket brigade are less so, but share that quality of fighting a combinatorial explosion, which results in painfully slow learning.
                 
                That's how I see things, at the level of hand waving argument. I don't know the literature well enough to point you at sources which would support these arguments in a more precise or rigorous way.
                ----- Original Message -----
                Sent: Tuesday, January 06, 2004 10:47 PM
                Subject: Re: [GP] Self Organization and scalable solutions with emergent phenomena

                David... please consider elaborating on your thoughts
                on why "one of the difficulties of credit assignment
                > algorithms such as back propagation in neural nets
                > and bucket brigade in classifier systems is that
                > they attempt a Lamarckian learning which is
                > inherently problematic."

                Would like to have this background as a 'heads-up' in
                looking into the workings of these other processes.

                Most finest regards,
                dennis



                __________________________________
                Do you Yahoo!?
                Yahoo! Hotjobs: Enter the "Signing Bonus" Sweepstakes
                http://hotjobssweepstakes.yahoo.com/signingbonus


                Yahoo! Groups Links

              Your message has been successfully submitted and would be delivered to recipients shortly.