Announcement

Collapse
No announcement yet.

Evolving Neural Networks with Genetic Algorithms for controlling Racing Cars.

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

    Evolving Neural Networks with Genetic Algorithms for controlling Racing Cars.

    Achievment unlocked: Evolving Neural Networks with Genetic Algorithms for controlling Racing Cars.

    #2
    can you provide source code to study ? or post it on github?

    Comment


      #3
      Ok, this completely blew my mind, well done good sir!

      Click the tiny bird!
      It'll take you to my stuff!

      Comment


        #4
        Originally posted by Xptohz3_turbo View Post
        Achievment unlocked: Evolving Neural Networks with Genetic Algorithms for controlling Racing Cars.
        Very awesome. Excellent work. Maximum overdrive —

        Comment


          #5
          Wow, this one got me!!

          Comment


            #6
            I did something similar, almost identical, in June 2016. Odd you used the same exact track.

            Did i inspire you ?

            Good job, though glad to see other people doing machine learning in UE4.
             
            Youtube
            Machine Learning C++ Plugin
            Lindenmayer System C++ Plugin

            Comment


              #7
              Awesome
              Is this made entirely inside UE4 or are there other thirdparty's ?
              Can you share more info about it?

              Comment


                #8
                I did something similar, but then evolving neural network controllers for arbitrary embodiments (in this case 4 legged).

                Enjoy!

                 

                Comment


                  #9
                  Originally posted by SaxonRah View Post
                  I did something similar, almost identical, in June 2016. Odd you used the same exact track.

                  Did i inspire you ?

                  Good job, though glad to see other people doing machine learning in UE4.
                  In fact, I saw your video and asked you questions. But got no reply.

                  I teach AI and I've been working with Genetic Algorithms for several years (I have a PhD on GAs).

                  I did this as a project assignment for my students.

                  I really enjoyed your video. Are the yellow cars NNs without Genetic Algorithms? How did you perform the training?

                  Thanks,
                  Gustavo

                  Comment


                    #10
                    Originally posted by TommieBommie View Post
                    I did something similar, but then evolving neural network controllers for arbitrary embodiments (in this case 4 legged).

                    Enjoy!
                    I already knew your work, and saw your video. Let me say that your work is amazing.

                    You did one of the things that I would like to accomplish. I would love to evolve neural network controllers for Biped and Quadruped and evolve the Idle, walk and run animations as in this video.

                     

                    Comment


                      #11
                      Originally posted by TommieBommie View Post
                      I did something similar, but then evolving neural network controllers for arbitrary embodiments (in this case 4 legged).

                      Enjoy!
                      I seen this a little while back and it blew me away. How difficult was it creating the creatures? Just some physic joints attached to some bodies?


                      Originally posted by Xptohz3_turbo View Post

                      In fact, I saw your video and asked you questions. But got no reply.

                      I teach AI and I've been working with Genetic Algorithms for several years (I have a PhD on GAs).

                      I did this as a project assignment for my students.

                      I really enjoyed your video. Are the yellow cars NNs without Genetic Algorithms? How did you perform the training?

                      Thanks,
                      Gustavo
                      D: I'm sorry! YT doesn't do the best of job of notifying me of things.

                      Oh that's totally awesome! I'm working on my PhD right now

                      I bet they can learn allot about AI from how you implemented the interface with the cars. That's super cool, i wish i had thought of that.

                      IIRC the Yellow cars are relatively simple avoidance networks. They were trained on data from how the Blue cars using BP Checkpoint AI were doing.

                      One thing of interest I noticed is that shock wave traffic is an emergent property of the Yellow Cars once they start off. Once they figure out a following distance the shock wave traffic is no longer apparent. Although, I would imagine if the track had more cars than the following distance would allow, or a larger following distance, (basically too many cars for the track) the shock wave traffic would emerge.

                      What I have concluded from that is, traffic is essentially human error / maliciousness and basically traffic could become a thing of the past if everyone drove the same.
                      Last edited by SaxonRah; 11-11-2017, 09:22 PM.
                      Youtube
                      Machine Learning C++ Plugin
                      Lindenmayer System C++ Plugin

                      Comment


                        #12
                        I saw it a little late, but this is fantastic, great job. I didn't even knew you can use UE for this.

                        Comment


                          #13
                          Originally posted by Xptohz3_turbo View Post

                          I already knew your work, and saw your video. Let me say that your work is amazing.

                          You did one of the things that I would like to accomplish. I would love to evolve neural network controllers for Biped and Quadruped and evolve the Idle, walk and run animations as in this video.
                          Yes the work shown in that video is really great and also inspired me to pick up on this and see if the Unreal Engine was up for the task. I actually also made an implementation with walking, but never got to implement a recurrent version of a neural network, instead I tried feeding in some sinusoidal functions but this didnt give the controller enough information to figure out a good behavior. Let me see if i still have some video material of this attempt, results were rather entertaining!

                          Found it!:

                          https://drive.google.com/file/d/19fB...ew?usp=sharing
                          Last edited by TommieBommie; 11-12-2017, 06:37 AM.

                          Comment


                            #14
                            Originally posted by SaxonRah View Post
                            I seen this a little while back and it blew me away. How difficult was it creating the creatures? Just some physic joints attached to some bodies?
                            The creatures are in deed just physic joints attached to some bodies. The NN drives the joints. It was however quite a bit of fiddling around to get the right settings, and also the Physics in Unreal Engine are not always precise enough (non deterministic i guess), i mitigated this problem by giving individuals multiple attempts to prove their fitness and averaging performance.

                            For my master project (Artificial Intelligence) I am planning some cool work on behavior trees and machine learning, i hope in the future to make a plugin available that allows all users of the UE4 to use this architecture, BUT first i need to graduate on the topic .

                            Comment


                              #15
                              Originally posted by TommieBommie View Post

                              The creatures are in deed just physic joints attached to some bodies. The NN drives the joints. It was however quite a bit of fiddling around to get the right settings, and also the Physics in Unreal Engine are not always precise enough (non deterministic i guess), i mitigated this problem by giving individuals multiple attempts to prove their fitness and averaging performance.

                              For my master project (Artificial Intelligence) I am planning some cool work on behavior trees and machine learning, i hope in the future to make a plugin available that allows all users of the UE4 to use this architecture, BUT first i need to graduate on the topic .
                              Nice! :-)

                              Good luck with your Masters!

                              Comment

                              Working...
                              X