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

Measuring the evolvability of HyperNEAT genomes

Expand Messages
  • martin_pyka
    Hi, inspired by the evolvability-article by Joel and Ken I was wondering how one would measure the degree of evolvability of an HyperNEAT encoding. I I would
    Message 1 of 4 , May 7, 2013
    • 0 Attachment
      Hi,

      inspired by the evolvability-article by Joel and Ken I was wondering how one would measure the degree of evolvability of an HyperNEAT encoding. I

      I would randomly change the weights of the NEAT network and measure the change in the phenotype. Which measure would be more appropriate, correlation coefficient or the euclidean distance in the n-dimensional phenotyp-space? Or are there better measures to assess evolvability in HyperNEAT?

      Best,
      Martin
    • Sebastian Risi
      Hi Martin, We actually compared the evolvability of HyperNEAT to Evolvable-Substrate HyperNEAT (ES-HyperNEAT) in our recent Artificial Life paper: Sebastian
      Message 2 of 4 , May 7, 2013
      • 0 Attachment
        Hi Martin,

        We actually compared the evolvability of HyperNEAT to
        Evolvable-Substrate HyperNEAT (ES-HyperNEAT) in our recent Artificial
        Life paper:

        Sebastian Risi and Kenneth O. Stanley (2012)
        An Enhanced Hypercube-Based Encoding for Evolving the Placement,
        Density and Connectivity of Neurons

        You can find the paper here: http://eplex.cs.ucf.edu/papers/risi_alife12.pdf

        Our definition of evolvability is similar to previous work from Joel
        and Ken and tries to quantify how well the underlying encoding enables
        behaviorally diverse mutations Figure 14).

        Best,

        Sebastian


        On Tue, May 7, 2013 at 8:44 AM, martin_pyka <martin.pyka@...> wrote:
        >
        >
        >
        > Hi,
        >
        > inspired by the evolvability-article by Joel and Ken I was wondering how one would measure the degree of evolvability of an HyperNEAT encoding. I
        >
        > I would randomly change the weights of the NEAT network and measure the change in the phenotype. Which measure would be more appropriate, correlation coefficient or the euclidean distance in the n-dimensional phenotyp-space? Or are there better measures to assess evolvability in HyperNEAT?
        >
        > Best,
        > Martin
        >
        >




        --
        Dr. Sebastian Risi
        Postdoctoral Fellow
        Creative Machines Laboratory
        Cornell University
        Email: sebastian.risi@... Tel: (407) 929-5113
        Web: http://www.cs.ucf.edu/~risi/
      • martin_pyka
        Hi Sebastian thanks for the link. But I am more interested in measuring evolvability in the substrate itself rather than the behaviour that is generated by a
        Message 3 of 4 , May 7, 2013
        • 0 Attachment
          Hi Sebastian

          thanks for the link. But I am more interested in measuring evolvability in the substrate itself rather than the behaviour that is generated by a neural network based on the substrate. Any ideas about that?

          Best,
          Martin

          --- In neat@yahoogroups.com, Sebastian Risi <sebastian.risi@...> wrote:
          >
          > Hi Martin,
          >
          > We actually compared the evolvability of HyperNEAT to
          > Evolvable-Substrate HyperNEAT (ES-HyperNEAT) in our recent Artificial
          > Life paper:
          >
          > Sebastian Risi and Kenneth O. Stanley (2012)
          > An Enhanced Hypercube-Based Encoding for Evolving the Placement,
          > Density and Connectivity of Neurons
          >
          > You can find the paper here: http://eplex.cs.ucf.edu/papers/risi_alife12.pdf
          >
          > Our definition of evolvability is similar to previous work from Joel
          > and Ken and tries to quantify how well the underlying encoding enables
          > behaviorally diverse mutations Figure 14).
          >
          > Best,
          >
          > Sebastian
          >
          >
          > On Tue, May 7, 2013 at 8:44 AM, martin_pyka <martin.pyka@...> wrote:
          > >
          > >
          > >
          > > Hi,
          > >
          > > inspired by the evolvability-article by Joel and Ken I was wondering how one would measure the degree of evolvability of an HyperNEAT encoding. I
          > >
          > > I would randomly change the weights of the NEAT network and measure the change in the phenotype. Which measure would be more appropriate, correlation coefficient or the euclidean distance in the n-dimensional phenotyp-space? Or are there better measures to assess evolvability in HyperNEAT?
          > >
          > > Best,
          > > Martin
          > >
          > >
          >
          >
          >
          >
          > --
          > Dr. Sebastian Risi
          > Postdoctoral Fellow
          > Creative Machines Laboratory
          > Cornell University
          > Email: sebastian.risi@... Tel: (407) 929-5113
          > Web: http://www.cs.ucf.edu/~risi/
          >
        • Ken
          Hi Martin, I think the challenge with looking at just the substrate is deciding how to determine what is novel just on the basis of the substrate. Technically
          Message 4 of 4 , May 7, 2013
          • 0 Attachment
            Hi Martin, I think the challenge with looking at just the substrate is deciding how to determine what is novel just on the basis of the substrate. Technically the problem isn't that hard because if you are using regular HyperNEAT (as opposed to ES-HyperNEAT) then all the substrates in a population have the same maximum number of weights, so they can be lined up on a connection-by-connection basis. However, conceptually the problem is hard because you are looking at nothing but weight patterns. You could measure average distance on weight-weight pairs, but that might not be meaningful enough to tell you much. It's a lot like measuring evolvability in Picbreeder - how would you decide if one image is sufficiently different from another to be considered a new kind of image? You might consider image matching algorithms, which could perhaps give a bit deeper insight than just pixel-by-pixel (or weight-by-weight) matching (except that in substrates you'd be matching images in 4D or 6D rather than 2D). Still, it's an interesting idea to try to measure in any case.

            Best,

            ken


            --- In neat@yahoogroups.com, "martin_pyka" <martin.pyka@...> wrote:
            >
            > Hi Sebastian
            >
            > thanks for the link. But I am more interested in measuring evolvability in the substrate itself rather than the behaviour that is generated by a neural network based on the substrate. Any ideas about that?
            >
            > Best,
            > Martin
            >
            > --- In neat@yahoogroups.com, Sebastian Risi <sebastian.risi@> wrote:
            > >
            > > Hi Martin,
            > >
            > > We actually compared the evolvability of HyperNEAT to
            > > Evolvable-Substrate HyperNEAT (ES-HyperNEAT) in our recent Artificial
            > > Life paper:
            > >
            > > Sebastian Risi and Kenneth O. Stanley (2012)
            > > An Enhanced Hypercube-Based Encoding for Evolving the Placement,
            > > Density and Connectivity of Neurons
            > >
            > > You can find the paper here: http://eplex.cs.ucf.edu/papers/risi_alife12.pdf
            > >
            > > Our definition of evolvability is similar to previous work from Joel
            > > and Ken and tries to quantify how well the underlying encoding enables
            > > behaviorally diverse mutations Figure 14).
            > >
            > > Best,
            > >
            > > Sebastian
            > >
            > >
            > > On Tue, May 7, 2013 at 8:44 AM, martin_pyka <martin.pyka@> wrote:
            > > >
            > > >
            > > >
            > > > Hi,
            > > >
            > > > inspired by the evolvability-article by Joel and Ken I was wondering how one would measure the degree of evolvability of an HyperNEAT encoding. I
            > > >
            > > > I would randomly change the weights of the NEAT network and measure the change in the phenotype. Which measure would be more appropriate, correlation coefficient or the euclidean distance in the n-dimensional phenotyp-space? Or are there better measures to assess evolvability in HyperNEAT?
            > > >
            > > > Best,
            > > > Martin
            > > >
            > > >
            > >
            > >
            > >
            > >
            > > --
            > > Dr. Sebastian Risi
            > > Postdoctoral Fellow
            > > Creative Machines Laboratory
            > > Cornell University
            > > Email: sebastian.risi@ Tel: (407) 929-5113
            > > Web: http://www.cs.ucf.edu/~risi/
            > >
            >
          Your message has been successfully submitted and would be delivered to recipients shortly.