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

Is someone working on a project using NEAT

Expand Messages
  • Mitchell Timin
    I ve seen very little reporting here of work in progress, nor releases, nor projects soon to be released. This surprises me. Surely I m not the only code
    Message 1 of 13 , Dec 1, 2003
    • 0 Attachment
      I've seen very little reporting here of work in
      progress, nor releases, nor projects soon to be
      released. This surprises me. Surely I'm not the only
      code monkey here. :) How are you guys spending your time?

      =====
      Mitchell Timin
      http://annevolve.sourceforge.net

      __________________________________
      Do you Yahoo!?
      Free Pop-Up Blocker - Get it now
      http://companion.yahoo.com/
    • Derek James
      ... I ve mentioned it before, but Philip Tucker and I are working on a new Java implementation of NEAT that we hope to opensource the first version of within
      Message 2 of 13 , Dec 2, 2003
      • 0 Attachment
        --- In neat@yahoogroups.com, Mitchell Timin <zenguyuno@y...> wrote:
        > I've seen very little reporting here of work in
        > progress, nor releases, nor projects soon to be
        > released. This surprises me. Surely I'm not the only
        > code monkey here. :) How are you guys spending your time?

        I've mentioned it before, but Philip Tucker and I are working on a
        new Java implementation of NEAT that we hope to opensource the first
        version of within the next month or so.

        Our primary focus is on board game domains. Initially, we're going
        to be running some experiments with Tic-Tac-Toe and Gomoku (five-in-a-
        row). Some of the areas we're interested in exploring are:

        1) Coevolutionary Algorithms -- I'd like to do some extensive
        comparisons of various coevolutionary algorithms, such as comparisons
        between single and two-or-more population coevolution, and the most
        effective ways to sample opponents each generation to ensure that
        desirable coevolutionary "arms races" continue throughout a run.

        2) Variations on NEAT -- The "A" stands for "augmenting", and all the
        topological mutations in NEAT are additive. We'd like to experiment
        with subtractive mutations, as a way of pruning inefficient or
        unnecessary topology in a mature individual.

        3) Indirect Encoding -- We've thought of some approaches to indirect
        encoding within the context of NEAT. I'd personally like to
        experiment with mutations that replace single neurons with modular
        subnets, kept in a pool (as in SANE and Joseph's modular NEAT
        experiments).

        So there are lots of different experimental tangents we'd like to
        explore in the coming year, though these are the three that interest
        me the most. But for the moment, we're still coding away...

        Derek
      • Kenneth Stanley
        ... I know a number of people in this group are working on NEAT-related projects. I think a lot of people are lurking, or mentioned their projects before and
        Message 3 of 13 , Dec 4, 2003
        • 0 Attachment
          --- In neat@yahoogroups.com, Mitchell Timin <zenguyuno@y...> wrote:
          > I've seen very little reporting here of work in
          > progress, nor releases, nor projects soon to be
          > released. This surprises me. Surely I'm not the only
          > code monkey here. :) How are you guys spending your time?
          >

          I know a number of people in this group are working on NEAT-related
          projects. I think a lot of people are lurking, or mentioned their
          projects before and don't feel like describing them again.

          So since this question comes up once a while, here's what I did: I
          added a bunch of projects I know of to the following part of the NEAT
          Users Page:

          http://www.cs.utexas.edu/users/kstanley/neat.html#users

          For the ones I just added, I did not include names or links in most
          cases, in order to protect the anonymity of the researcher/hobbyist
          involved. So, if you see your project listed there, and you want to
          get credit or have it linked, let me know and I will add in your name
          and link. (Likewise if you want your project removed from the list ,
          I will remove it at your request)

          Hope this helps answer your question.

          ken
        • Mitchell Timin
          ... ... http://www.cs.utexas.edu/users/kstanley/neat.html#users As a result of going there and clicking around a bit, I just ordered Mat Buckland s
          Message 4 of 13 , Dec 4, 2003
          • 0 Attachment
            --- Kenneth Stanley <kstanley@...> wrote:

            <snip>

            > So since this question comes up once a while, here's
            > what I did: I
            > added a bunch of projects I know of to the following
            > part of the NEAT
            > Users Page:

            http://www.cs.utexas.edu/users/kstanley/neat.html#users

            As a result of going there and clicking around a bit,
            I just ordered Mat Buckland's book. :)

            <snip>

            > Hope this helps answer your question.

            Yes, Thanks.


            =====
            Mitchell Timin
            http://annevolve.sourceforge.net

            __________________________________
            Do you Yahoo!?
            Free Pop-Up Blocker - Get it now
            http://companion.yahoo.com/
          • Roger Smith
            I was thinking of using NEAT in the JRobots competition. Does anyone have any advice for me here? Thanks Rog __________________________________ Do you Yahoo!?
            Message 5 of 13 , Dec 6, 2003
            • 0 Attachment
              I was thinking of using NEAT in the JRobots
              competition.
              Does anyone have any advice for me here?
              Thanks
              Rog



              __________________________________
              Do you Yahoo!?
              New Yahoo! Photos - easier uploading and sharing.
              http://photos.yahoo.com/
            • Kenneth Stanley
              ... Hey Rog, could you repost the link to JRobots? thanks, ken
              Message 6 of 13 , Dec 7, 2003
              • 0 Attachment
                --- In neat@yahoogroups.com, Roger Smith <rog_21@y...> wrote:
                > I was thinking of using NEAT in the JRobots
                > competition.
                > Does anyone have any advice for me here?
                > Thanks
                > Rog
                >

                Hey Rog, could you repost the link to JRobots?

                thanks,
                ken
              • Roger Smith
                http://www.cfxweb.net/~jrobots/ ... __________________________________ Do you Yahoo!? New Yahoo! Photos - easier uploading and sharing.
                Message 7 of 13 , Dec 7, 2003
                • 0 Attachment
                  http://www.cfxweb.net/~jrobots/


                  --- Kenneth Stanley <kstanley@...> wrote:
                  > --- In neat@yahoogroups.com, Roger Smith
                  > <rog_21@y...> wrote:
                  > > I was thinking of using NEAT in the JRobots
                  > > competition.
                  > > Does anyone have any advice for me here?
                  > > Thanks
                  > > Rog
                  > >
                  >
                  > Hey Rog, could you repost the link to JRobots?
                  >
                  > thanks,
                  > ken
                  >
                  >


                  __________________________________
                  Do you Yahoo!?
                  New Yahoo! Photos - easier uploading and sharing.
                  http://photos.yahoo.com/
                • Derek James
                  ... Hey Roger, Well, I m certainly interested in how you could design a bot for JRobots using NEAT, but I ve mostly been focusing on working on our NEAT
                  Message 8 of 13 , Dec 8, 2003
                  • 0 Attachment
                    --- In neat@yahoogroups.com, Roger Smith <rog_21@y...> wrote:
                    > I was thinking of using NEAT in the JRobots
                    > competition.
                    > Does anyone have any advice for me here?

                    Hey Roger,

                    Well, I'm certainly interested in how you could design a bot for
                    JRobots using NEAT, but I've mostly been focusing on working on our
                    NEAT implementation and prepping for experiments with board game
                    domains.

                    Their documentation didn't look entirely user-friendly, though.
                    Sorry I can't be of help, but be sure to keep us posted on your
                    efforts if you are able to make any headway.

                    Derek
                  • Derek James
                    This is mostly addressed to Ken, but I thought I d also open it to the group. I just finished reading de Jong and Pollack s yet-to-be-published paper Ideal
                    Message 9 of 13 , Dec 8, 2003
                    • 0 Attachment
                      This is mostly addressed to Ken, but I thought I'd
                      also open it to the group.

                      I just finished reading de Jong and Pollack's
                      yet-to-be-published paper "Ideal Evaluation from
                      Coevolution" via de Jong's website:

                      http://www.cs.uu.nl/people/dejong/publications/coevec.pdf

                      I noticed that you were among those thanked at the end
                      of the paper, so I assume you reviewed it. What did
                      you think of it?

                      I found some of the ideas interesting, though I didn't
                      understand some, and I wasn't especially clear on the
                      experimental setup.

                      I like the idea of Multi-Objective Optimization and of
                      specifying a population of Evaluators that test
                      Learners on specific objectives, though I don't admit
                      to understanding how this actually works in practice.

                      I've already found in some informal experimentation
                      with unspeciated NEAT applied to Tic-Tac-Toe that
                      populations evolved directly against random movers
                      learn how to complete 2-in-a-row to get to a
                      3-in-a-row win, and that those evolved directly
                      against perfect players learn how to block opponent's
                      winning moves but perform poorly against random movers
                      because they never learn how to complete winning moves
                      (a perfect player never allows this situation).

                      Can these two behavioral traits (completing winning
                      moves and blocking opponent's winning moves) be viewed
                      as discrete objectives? And then both of these types
                      of Evaluators (one that allows completions of winning
                      moves and one that makes winning moves to be blocked)
                      would have to evolve in de Jong and Pollack's DELPHI
                      paradigm, in order to test Learners against?

                      I'm just not clear on how Evaluators are chosen to be
                      representatives of specific objectives and how DELPHI
                      would work in practice with a game domain, even a
                      simple one such as Tic-Tac-Toe.

                      Thanks,
                      Derek

                      __________________________________
                      Do you Yahoo!?
                      Protect your identity with Yahoo! Mail AddressGuard
                      http://antispam.yahoo.com/whatsnewfree
                    • Kenneth Stanley
                      ... Hey Rog, JRobots is definitely the type of thing NEAT would be good at. The first thing you need to decide is which version of NEAT to use, and it looks
                      Message 10 of 13 , Dec 8, 2003
                      • 0 Attachment
                        --- In neat@yahoogroups.com, Roger Smith <rog_21@y...> wrote:
                        > I was thinking of using NEAT in the JRobots
                        > competition.
                        > Does anyone have any advice for me here?
                        > Thanks
                        > Rog
                        >
                        >
                        >

                        Hey Rog, JRobots is definitely the type of thing NEAT
                        would be good at. The first thing you need to decide
                        is which version of NEAT to use, and it looks like
                        JNEAT may be the best since JRobots wants a robot coded
                        in Java. However, you may also want to talk to Derek
                        and Philip about their release shedule of their Java
                        version.

                        Next, you need to figure out what the NN inputs and outputs
                        should be. You need to decide how to translate the
                        input from the game into a nice(preferably small)
                        set of NN inputs, and also how the outputs from the NN
                        should map to the control signals in the game.

                        Once you have that down, you are ready to start evolution
                        in their simulator. The hardest part will be integrating
                        whatever version of NEAT you choose into their simulator.
                        However, it is well worth the effort since this is
                        certainly a good application of NEAT.

                        ken
                      • Kenneth Stanley
                        ... Yes I did read it, though I wasn t an official reviewer. In general, the official reviewers are anonymous. But I read it since I know Edwin de Jong and
                        Message 11 of 13 , Dec 11, 2003
                        • 0 Attachment
                          --- In neat@yahoogroups.com, Derek James <blue5432@y...> wrote:
                          > This is mostly addressed to Ken, but I thought I'd
                          > also open it to the group.
                          >
                          > I just finished reading de Jong and Pollack's
                          > yet-to-be-published paper "Ideal Evaluation from
                          > Coevolution" via de Jong's website:
                          >
                          > http://www.cs.uu.nl/people/dejong/publications/coevec.pdf
                          >
                          > I noticed that you were among those thanked at the end
                          > of the paper, so I assume you reviewed it. What did
                          > you think of it?
                          >

                          Yes I did read it, though I wasn't an official reviewer. In general,
                          the official reviewers are anonymous. But I read it since I know
                          Edwin de Jong and he sent me a copy, and I sent him back a lot of
                          comments. I thought it was a very good paper. It really gives you a
                          broad general sense of what it takes to become good at a particular
                          game, and a theoretical sense of how challenging that really is.
                          Because of this, I am not certain that the current form of the
                          algorithm is really that practical, since it is based on creating a
                          theoretically optimal competition. The problem is that in practice,
                          it is probably too computationally expensive to do that. (That is,
                          having all the players play all the test cases would just take too
                          long in any serious game.) I spoke briefly to de Jong about this and
                          he agreed at the time that some kind of sampling (i.e. not doing all
                          the comparisons) might be necessary in some domains, and I kind of
                          have the feeling that the "next generation" of this algorithm would
                          include a principled sampling component. That, I believe, would make
                          a really good competitive coevolution methodology, and probably would
                          replace what I consider the current best practical method (Rosin and
                          Belew 1997).

                          > I found some of the ideas interesting, though I didn't
                          > understand some, and I wasn't especially clear on the
                          > experimental setup.
                          >
                          > I like the idea of Multi-Objective Optimization and of
                          > specifying a population of Evaluators that test
                          > Learners on specific objectives, though I don't admit
                          > to understanding how this actually works in practice.
                          >
                          > I've already found in some informal experimentation
                          > with unspeciated NEAT applied to Tic-Tac-Toe that
                          > populations evolved directly against random movers
                          > learn how to complete 2-in-a-row to get to a
                          > 3-in-a-row win, and that those evolved directly
                          > against perfect players learn how to block opponent's
                          > winning moves but perform poorly against random movers
                          > because they never learn how to complete winning moves
                          > (a perfect player never allows this situation).
                          >
                          > Can these two behavioral traits (completing winning
                          > moves and blocking opponent's winning moves) be viewed
                          > as discrete objectives? And then both of these types
                          > of Evaluators (one that allows completions of winning
                          > moves and one that makes winning moves to be blocked)
                          > would have to evolve in de Jong and Pollack's DELPHI
                          > paradigm, in order to test Learners against?
                          >

                          Yes I think you make a good conceptual anology. Your two ways of
                          evaluating tic tac toe are roughly two objectives. What would happen
                          in practice with the DELPHI algorithm is that a whole lot other
                          objectives would be revealed and you would be able to construct a
                          "front" that represented that various ways that players could be good.

                          > I'm just not clear on how Evaluators are chosen to be
                          > representatives of specific objectives and how DELPHI
                          > would work in practice with a game domain, even a
                          > simple one such as Tic-Tac-Toe.
                          >

                          My guess would be it actually could work well with Tic-Tac-Toe, since
                          Tic-Tac-Toe is such a computationally inexpensive game. The
                          computational power is probably there to totally apply the DELPHI
                          methodology and get really good results. Of course all this assumes
                          that the other parts of your evolutionary algorithm work well too.

                          ken
                        • Mitchell Timin
                          This topic was mentioned a couple of time in the recent thread. I just wanted to point out that natural biological evolution is directed by numerous
                          Message 12 of 13 , Dec 11, 2003
                          • 0 Attachment
                            This topic was mentioned a couple of time in the
                            recent thread. I just wanted to point out that
                            natural biological evolution is directed by numerous
                            objectives, too numerous to enumerate! Broadly, an
                            organism seeks food, safety from predators, safety
                            from micro-pathogens, shelter, thermo-regulation,
                            recovery from damage or illness, reproductive
                            opportunity, strategy for maturation of young, and so
                            on. Each of those in turn has complications within
                            it.
                            So we may conclude that evolution is a process that
                            can well handle multiple objectives.

                            =====
                            Mitchell Timin
                            http://annevolve.sourceforge.net

                            __________________________________
                            Do you Yahoo!?
                            Free Pop-Up Blocker - Get it now
                            http://companion.yahoo.com/
                          • Derek James
                            ... Ah...okay. ... Yeah. :) ... Yes, I can see how that would be a problem. How to select the most effective and efficient sample of opponents in a
                            Message 13 of 13 , Dec 12, 2003
                            • 0 Attachment
                              --- Kenneth Stanley <kstanley@...> wrote:
                              > Yes I did read it, though I wasn't an official
                              > reviewer. In general,
                              > the official reviewers are anonymous.

                              Ah...okay.

                              > I thought it was a very good paper. It
                              > really gives you a
                              > broad general sense of what it takes to become good
                              > at a particular
                              > game, and a theoretical sense of how challenging
                              > that really is.

                              Yeah. :)

                              > Because of this, I am not certain that the current
                              > form of the
                              > algorithm is really that practical, since it is
                              > based on creating a
                              > theoretically optimal competition. The problem is
                              > that in practice,
                              > it is probably too computationally expensive to do
                              > that. (That is,
                              > having all the players play all the test cases would
                              > just take too
                              > long in any serious game.)

                              Yes, I can see how that would be a problem. How to
                              select the most effective and efficient sample of
                              opponents in a coevolutionary algorithm seems like an
                              especially difficult problem.

                              I like the systematic approach of DELPHI, but I
                              wondered just how it would work in practice on
                              something other than a toy domain. Seems like you're
                              saying it wouldn't.

                              > That, I
                              > believe, would make
                              > a really good competitive coevolution methodology,
                              > and probably would
                              > replace what I consider the current best practical
                              > method (Rosin and
                              > Belew 1997).

                              Yeah, I'm really fascinated with this topic. Are
                              there any other resources you (or anybody else here)
                              might point me to?

                              It seems like the key is to find an efficient sample
                              of opponents each generation that tests each
                              individual against a diverse range of objectives. I
                              believe it was one of your robot domains in which you
                              sampled opponents from across each NEAT species and
                              used dominance tournaments to ensure a strong, diverse
                              opponent population each generation.

                              I've read some of the discussion of "subjective" and
                              "objective" fitness evaluations in coevolutionary
                              algorithms (e.g. Pollack, Luke, etc.). It seems to me
                              as if there's almost always an either/or with regard
                              to such experiments. That is, individuals are either
                              evolved against internal opponents (others in the same
                              population or past champions from the same run) or
                              external opponents (pre-programmed algorithms). To
                              your knowledge, have there been any experiments with a
                              blend of subjective and objective evaluations? I'm
                              curious to see how this would play out experimentally.

                              > Yes I think you make a good conceptual anology.
                              > Your two ways of
                              > evaluating tic tac toe are roughly two objectives.
                              > What would happen
                              > in practice with the DELPHI algorithm is that a
                              > whole lot other
                              > objectives would be revealed and you would be able
                              > to construct a
                              > "front" that represented that various ways that
                              > players could be good.

                              This is an interesting concept. I mean, a single
                              player could represent *all* the ways that various
                              players could be good (this optimal strategy is what
                              we want to evolve toward). But players evolved
                              against such a player would not be very robust.

                              Are the ways that given strategies can be decomposed
                              and represented by individual players dependent mostly
                              on the domain? I mean, using the DELPHI technique,
                              you think you would end up with a given number of
                              players, say 20 or 30, that would each represent a
                              test of a given strategic objective?

                              To what extent are strategies decomposable? Could a
                              given individual test for two or more objectives?

                              Derek

                              __________________________________
                              Do you Yahoo!?
                              Protect your identity with Yahoo! Mail AddressGuard
                              http://antispam.yahoo.com/whatsnewfree
                            Your message has been successfully submitted and would be delivered to recipients shortly.