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

Re: [GP] Has anyone tested Grammatical Evolution?

Expand Messages
  • R. Muhammad Atif Azad
    Hi, GE is generally known to be a system of plugins and therefore has a number of design parameters that add to its flexibility. These include the search
    Message 1 of 7 , May 24, 2005
      Hi,
      GE is generally known to be a system of plugins and therefore has a number
      of design parameters that add to its flexibility. These include the search
      algorithm which is traditionally a simple genetic algorithm but any other
      algorithm that works at bit strings can be used (however, different
      algorithms will produce different results). Likewise grammar is also
      an important design choice and I suspect any grammar guided algorithm will
      be sensitive to the grammar its seeded with.

      Moreover, what is the comparison measure you are using? If it is the
      cumulative frequency of success, questions have been raised lately over
      its reliability such as 'Is the perfect the enemy of the good' by Sean
      Luke and L. Paniat in GECCO 2002. My experience with GE is that in terms
      of mean best fitness it does pretty well on the quartic polynomial
      problem.

      Moreover, in an article to appear in 'Genetic Programming Theory and
      Practice 2005' we observed that a single non-terminal
      grammar seems to be the most suitable. It may be so because many of the
      GP benchmarks such as Quartic polynomial, parity and multiplexer problems
      and santa fe ant trail do not require multiple types and context shifts
      arising due to multiple non-terminals may be an unnecessary imposition.

      Also we proposed that it may be more useful to restrict the crossover
      point to the mapping-effective lengths of the GE individuals. It is so
      because GE grows 'tails' over time i.e. genetic material beyond the length
      of the genome required to map an individual to completion. Long tails
      will cause crossover to swap the material mostly from the tails, thereby
      leaving the mapping-effective genetic material intact. Thus, the
      offsprings will tend to duplicate the phenotypes of their parents.

      See if playing with these ideas helps.
      Regards,
      Atif


      --
      Dr. Raja Muhammad Atif Azad
      CSIS Department
      University of Limerick
      Limerick
      Ireland.
      Voice 353-61-202763
      Fax 353-61-202734

      On Tue, 24 May 2005, Denis Robilliard wrote:

      > In my experience, GE is very sensitive to the grammar : a simple swap of
      > two alternative derivation rules may have a devastating effect, notably
      > due to wrapping (and wrapping is useful: don't remove it). GE may also
      > have difficulties with problems more complex than the simple toy
      > examples used in most of the litterature.
      >
      > Denis Robilliard
      >
      > winplusx0 wrote:
      >
      >> I tried Symbolic Regression problem and found the result is not as good
      >> as mentioned in published paper. Acturely it is very poor compared with
      >> traditional GP.
      >>
      >> I dont know what's wrong. Or my program's parameters are not correct?(
      >> I used most common configuation: steady-state,one bit crossover(0.9)...)
      >>
      >> Please help me.
      >>
      >>
      >>
      >>
      >>
      >>
      >>
      >>
      >> Yahoo! Groups Links
      >>
      >>
      >>
      >>
      >>
      >>
      >>
      >>
      >
      >
      >
      >
      > Yahoo! Groups Links
      >
      >
      >
      >
      >
      >
    • Michael Korns
      Often, in GE experiments when things just don t work at all, it is helpful to validate all floating point numbers. In symbolic regression and other numeric GE
      Message 2 of 7 , May 25, 2005
        Often, in GE experiments when things just don't work at all, it is helpful
        to validate all floating point numbers.

        In symbolic regression and other numeric GE experiments, many individuals GE
        generates are basically flawed and output invalid or out of range IEEE
        doubles. If left unchecked, these values will cause serious problems with
        the experimental results.

        Try validating all IEEE doubles and reject any individuals which return
        invalid numeric results.

        >I tried Symbolic Regression problem and found the result is not as good
        >as mentioned in published paper. Acturely it is very poor compared with
        >traditional GP.
        >
        >I dont know what's wrong. Or my program's parameters are not correct?(
        >I used most common configuation: steady-state,one bit crossover(0.9)...)
        >
        >Please help me.

        Michael

        ************************************************
        Michael Korns
        1 Plum Hollow Drive
        Henderson, Nevada 89052
        (702) 837-3498
        mkorns@...
        www.korns.com
        www.InvestByAgent.com
        ************************************************
      • winplusx0
        I have tried validate all floating point numbers and it doesnot work. In steady state algorithm, the bit flip mutation operator has two probability value:
        Message 3 of 7 , May 31, 2005
          I have tried validate all floating point numbers and it doesnot work.

          In steady state algorithm, the bit flip mutation operator has two
          probability value: individual mutation prob. and bit mutation prob.
          How to set these two?
          I use beagle as my evolution algorithm library.

          --- In genetic_programming@yahoogroups.com, "Michael Korns"
          <mkorns@k...> wrote:
          > Often, in GE experiments when things just don't work at all, it is
          helpful
          > to validate all floating point numbers.
          >
          > In symbolic regression and other numeric GE experiments, many
          individuals GE
          > generates are basically flawed and output invalid or out of range
          IEEE
          > doubles. If left unchecked, these values will cause serious
          problems with
          > the experimental results.
          >
          > Try validating all IEEE doubles and reject any individuals which
          return
          > invalid numeric results.
          >
          > >I tried Symbolic Regression problem and found the result is not
          as good
          > >as mentioned in published paper. Acturely it is very poor
          compared with
          > >traditional GP.
          > >
          > >I dont know what's wrong. Or my program's parameters are not
          correct?(
          > >I used most common configuation: steady-state,one bit crossover
          (0.9)...)
          > >
          > >Please help me.
          >
          > Michael
          >
          > ************************************************
          > Michael Korns
          > 1 Plum Hollow Drive
          > Henderson, Nevada 89052
          > (702) 837-3498
          > mkorns@k...
          > www.korns.com
          > www.InvestByAgent.com
          > ************************************************
        • winplusx0
          Thanks a lot. I have developed a new operator named Effectvie Crossover as mentioned in your reply which selects muting point in effective lengths of the GE
          Message 4 of 7 , Jun 3, 2005
            Thanks a lot.

            I have developed a new operator named "Effectvie Crossover" as
            mentioned in your reply which selects muting point in effective
            lengths of the GE genotype. It indeed improves the performance.But
            it still cannot reach the result mentioned in [1](Several other
            papers of the authors also have the experiment result).

            My best result with Symbolic Regression problem (X*X*X*X+X*X*X+X*X+X)
            is about 60% cumulative frequency of success in 50
            generations.However the paper obove mentioned a above 90% success in
            20 generations. My BNF for the problem is the same as in [1].

            BTW, I cannot visit www.grammatical-evolution.org recently. Does
            anyone know the reason?

            [1]O¡¯Neill, M., Ryan, C. (2001) Grammatical Evolution, IEEE Trans.
            Evolutionary Computation, 5(4):349-358, 2001.

            --- In genetic_programming@yahoogroups.com, "R. Muhammad Atif Azad"
            <atif.azad@u...> wrote:
            > Hi,
            > GE is generally known to be a system of plugins and therefore has
            a number
            > of design parameters that add to its flexibility. These include
            the search
            > algorithm which is traditionally a simple genetic algorithm but
            any other
            > algorithm that works at bit strings can be used (however,
            different
            > algorithms will produce different results). Likewise grammar is
            also
            > an important design choice and I suspect any grammar guided
            algorithm will
            > be sensitive to the grammar its seeded with.
            >
            > Moreover, what is the comparison measure you are using? If it is
            the
            > cumulative frequency of success, questions have been raised lately
            over
            > its reliability such as 'Is the perfect the enemy of the good' by
            Sean
            > Luke and L. Paniat in GECCO 2002. My experience with GE is that in
            terms
            > of mean best fitness it does pretty well on the quartic polynomial
            > problem.
            >
            > Moreover, in an article to appear in 'Genetic Programming Theory
            and
            > Practice 2005' we observed that a single non-terminal
            > grammar seems to be the most suitable. It may be so because many
            of the
            > GP benchmarks such as Quartic polynomial, parity and multiplexer
            problems
            > and santa fe ant trail do not require multiple types and context
            shifts
            > arising due to multiple non-terminals may be an unnecessary
            imposition.
            >
            > Also we proposed that it may be more useful to restrict the
            crossover
            > point to the mapping-effective lengths of the GE individuals. It
            is so
            > because GE grows 'tails' over time i.e. genetic material beyond
            the length
            > of the genome required to map an individual to completion. Long
            tails
            > will cause crossover to swap the material mostly from the tails,
            thereby
            > leaving the mapping-effective genetic material intact. Thus, the
            > offsprings will tend to duplicate the phenotypes of their parents.
            >
            > See if playing with these ideas helps.
            > Regards,
            > Atif
            >
            >
            > --
            > Dr. Raja Muhammad Atif Azad
            > CSIS Department
            > University of Limerick
            > Limerick
            > Ireland.
            > Voice 353-61-202763
            > Fax 353-61-202734
            >
            > On Tue, 24 May 2005, Denis Robilliard wrote:
            >
            > > In my experience, GE is very sensitive to the grammar : a simple
            swap of
            > > two alternative derivation rules may have a devastating effect,
            notably
            > > due to wrapping (and wrapping is useful: don't remove it). GE
            may also
            > > have difficulties with problems more complex than the simple toy
            > > examples used in most of the litterature.
            > >
            > > Denis Robilliard
            > >
            > > winplusx0 wrote:
            > >
            > >> I tried Symbolic Regression problem and found the result is not
            as good
            > >> as mentioned in published paper. Acturely it is very poor
            compared with
            > >> traditional GP.
            > >>
            > >> I dont know what's wrong. Or my program's parameters are not
            correct?(
            > >> I used most common configuation: steady-state,one bit crossover
            (0.9)...)
            > >>
            > >> Please help me.
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >> Yahoo! Groups Links
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >>
            > >
            > >
            > >
            > >
            > > Yahoo! Groups Links
            > >
            > >
            > >
            > >
            > >
            > >
          Your message has been successfully submitted and would be delivered to recipients shortly.