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

fitness algorithm for seek and avoid

Expand Messages
  • aaandy.doubleaa
    Hello everybody, this is my first post and I m looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with
    Message 1 of 5 , Feb 5, 2011
      Hello everybody,
      this is my first post and I'm looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with dynamic objects all implemented within the NEAT methodology, is there a de-facto algorithm already defined ? Or do I have to experiment myself?

      Cheers Andy
    • petar_chervenski
      Try a composite fitness function where the distance between the target has to be minimized (for each time step the fitness is MAX_DISTANCE - distance) and
      Message 2 of 5 , Feb 5, 2011
        Try a composite fitness function where the distance between the target has to be minimized (for each time step the fitness is MAX_DISTANCE - distance) and collisions with obstacles penalized (subtract a fixed value from the fitness score every time a collision occurs). Add these two and it should work fine, but there are many ways to do it, of course.

        Peter


        --- In neat@yahoogroups.com, "aaandy.doubleaa" <aaandy.doubleaa@...> wrote:
        >
        > Hello everybody,
        > this is my first post and I'm looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with dynamic objects all implemented within the NEAT methodology, is there a de-facto algorithm already defined ? Or do I have to experiment myself?
        >
        > Cheers Andy
        >
      • Ken
        You could also weight the importance of the two components to emphasize one over the other. However, in answer to Andy, there is no de facto correct way to
        Message 3 of 5 , Feb 8, 2011
          You could also weight the importance of the two components to emphasize one over the other. However, in answer to Andy, there is no de facto correct way to write the fitness function; there are several possibilities.


          ken

          --- In neat@yahoogroups.com, "petar_chervenski" <petar_chervenski@...> wrote:
          >
          > Try a composite fitness function where the distance between the target has to be minimized (for each time step the fitness is MAX_DISTANCE - distance) and collisions with obstacles penalized (subtract a fixed value from the fitness score every time a collision occurs). Add these two and it should work fine, but there are many ways to do it, of course.
          >
          > Peter
          >
          >
          > --- In neat@yahoogroups.com, "aaandy.doubleaa" <aaandy.doubleaa@> wrote:
          > >
          > > Hello everybody,
          > > this is my first post and I'm looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with dynamic objects all implemented within the NEAT methodology, is there a de-facto algorithm already defined ? Or do I have to experiment myself?
          > >
          > > Cheers Andy
          > >
          >
        • aaandy.doubleaa
          Dear Peter and Ken, Many thanks for the answer, I have been experimenting and I have tried fitness = s • (c1 - a ) • (c2 - m) • (c3/d) where s The
          Message 4 of 5 , Feb 18, 2011
            Dear Peter and Ken,
            Many thanks for the answer, I have been experimenting and I have tried
            fitness = s • (c1 - \a\) • (c2 - m) • (c3/d)
            where s The variable S is the speed of the neural agent m/sec
            a is Angular velocity Rads/sec(average )
            m = max sensor value currently detected by a sensor.
            d is distance to the goal

            c1,c2 and c3 are constants that I can play with to change emphasize of each factor currently c1 is 1 , c2 max sensor plus noise and c3 1 for punishment a small distance to the goal

            works well but a lot of spinning of the robot

            Many regards Andy
            --- In neat@yahoogroups.com, "Ken" <kstanley@...> wrote:
            >
            > You could also weight the importance of the two components to emphasize one over the other. However, in answer to Andy, there is no de facto correct way to write the fitness function; there are several possibilities.
            >
            >
            > ken
            >
            > --- In neat@yahoogroups.com, "petar_chervenski" <petar_chervenski@> wrote:
            > >
            > > Try a composite fitness function where the distance between the target has to be minimized (for each time step the fitness is MAX_DISTANCE - distance) and collisions with obstacles penalized (subtract a fixed value from the fitness score every time a collision occurs). Add these two and it should work fine, but there are many ways to do it, of course.
            > >
            > > Peter
            > >
            > >
            > > --- In neat@yahoogroups.com, "aaandy.doubleaa" <aaandy.doubleaa@> wrote:
            > > >
            > > > Hello everybody,
            > > > this is my first post and I'm looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with dynamic objects all implemented within the NEAT methodology, is there a de-facto algorithm already defined ? Or do I have to experiment myself?
            > > >
            > > > Cheers Andy
            > > >
            > >
            >
          • Ken
            Andy, to deal with spinning one thing you may want to try is making spinning physically impossible. That is, if the controller requests the same turn
            Message 5 of 5 , Feb 19, 2011
              Andy, to deal with spinning one thing you may want to try is making spinning physically impossible. That is, if the controller requests the same turn direction more than n times in row, simply refuse the request and don't move the robot. The result will be occasional frozen robots. However, they will evolve to deal with this problem, but of course, they cannot use spinning anymore, so whatever they come up with will not involve spinning.

              ken

              --- In neat@yahoogroups.com, "aaandy.doubleaa" <aaandy.doubleaa@...> wrote:
              >
              > Dear Peter and Ken,
              > Many thanks for the answer, I have been experimenting and I have tried
              > fitness = s • (c1 - \a\) • (c2 - m) • (c3/d)
              > where s The variable S is the speed of the neural agent m/sec
              > a is Angular velocity Rads/sec(average )
              > m = max sensor value currently detected by a sensor.
              > d is distance to the goal
              >
              > c1,c2 and c3 are constants that I can play with to change emphasize of each factor currently c1 is 1 , c2 max sensor plus noise and c3 1 for punishment a small distance to the goal
              >
              > works well but a lot of spinning of the robot
              >
              > Many regards Andy
              > --- In neat@yahoogroups.com, "Ken" <kstanley@> wrote:
              > >
              > > You could also weight the importance of the two components to emphasize one over the other. However, in answer to Andy, there is no de facto correct way to write the fitness function; there are several possibilities.
              > >
              > >
              > > ken
              > >
              > > --- In neat@yahoogroups.com, "petar_chervenski" <petar_chervenski@> wrote:
              > > >
              > > > Try a composite fitness function where the distance between the target has to be minimized (for each time step the fitness is MAX_DISTANCE - distance) and collisions with obstacles penalized (subtract a fixed value from the fitness score every time a collision occurs). Add these two and it should work fine, but there are many ways to do it, of course.
              > > >
              > > > Peter
              > > >
              > > >
              > > > --- In neat@yahoogroups.com, "aaandy.doubleaa" <aaandy.doubleaa@> wrote:
              > > > >
              > > > > Hello everybody,
              > > > > this is my first post and I'm looking for a fitness algorithm which can be used in a goal seek and obstacle avoidance within a maze with dynamic objects all implemented within the NEAT methodology, is there a de-facto algorithm already defined ? Or do I have to experiment myself?
              > > > >
              > > > > Cheers Andy
              > > > >
              > > >
              > >
              >
            Your message has been successfully submitted and would be delivered to recipients shortly.