Skip navigation
As I like cycling, genetic algorithms and interactive visualisations, I've made a neuroevolution…
0:00 / 0:26

As I like cycling, genetic algorithms and interactive visualisations, I've made a neuroevolution…

Andrew Davison

@AjdDavison22K followers

Visit on X
561 likesDownloadOriginal post

52,926 views

As I like cycling, genetic algorithms and interactive visualisations, I've made a neuroevolution sim. Riders are controlled by small random neural networks which evolve race by race and we see nice tactics emerge. Runs in the browser, best on a computer:

35 Comments

Comments from the original post on X (Twitter)

  • Andrew Davison's profile picture

    @AjdDavison

    Leave it running for half a hour in the background and see what happens. It has fairly accurate cycling physics, including drafting. Let me know any ideas to make it better!

    5 likes

  • Gill's profile picture

    @gurtej__gill_

    What signals does each neural network get as input during the race?

    4 likes

  • Andrew Davison's profile picture

    @AjdDavison

    It's explained in the text underneath.

    1 like

  • Ki's profile picture

    @Ki_fun_thoughts

    I co-created the CompuTrainer system. tl;dr - I once mistook Lance Armstrong's profile data for a computer-generated Pacer. You can 'see' doping. This is what mine looked like decades ago. AMA :)

    3 likes

  • Andrew Davison's profile picture

    @AjdDavison

    That's awesome

    1 like

  • Ki's profile picture

    @Ki_fun_thoughts

    Truly, where one can 'see' something not quite right, was in the hysterics. It would average out just slightly higher, but where most would fatigue with a bump while ascending, he did not. Topic change - helmet or no helmet? :)

  • Andrew Davison's profile picture

    @AjdDavison

    They should be wearing helmets at that speed for sure... the nicest sprite I could find though didn't have them! Don't try this at home kids.

    1 like

  • Ki's profile picture

    @Ki_fun_thoughts

    Indeed, I suspect most humans should be wearing helmets, physically or metaphysically, and at any speed. And if not for themselves, to reduce the damage they do to others, sort of like an N95 mask :) Ride through life safely.

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    Hi Andrew, I managed to get to a reasonable state with a similar version(I'm using the neat pkg here), and here is the demo (not as sleek as yours). Added a few things I thought might be worthwhile to help debug and build intuition there. (it's open source, I've cited your work)

    1 like

  • Andrew Davison's profile picture

    @AjdDavison

    Wow, nice work! I will take a proper look!

    1 like

  • Andrew Davison's profile picture

    @AjdDavison

    I don't know about the NEAT algorithm. How does that work?

    1 like

  • Andrew Davison's profile picture

    @AjdDavison

    I assume you are also a cycling fan!

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    Yes, we have a few tracks closer where I live.

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    Hi again Andrew, I looked into some of your recent work, i got really fascinated by the (robot web) this talk: so I started building my own simulation by going through your paper, and now I'm thinking why not, evalutionry robot web, genetic layer that monitors the position estimation, i've not started that execrise, and it would be awesome to collect your thoughts here

    1 like

  • Andrew Davison's profile picture

    @AjdDavison

    Glad you found that. It's my favourite paper. Can you explain more what you mean about adding evolution?

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    If the sensor noise is well-behaved, I think the GBP is already optimal. But what if the environment is non-convex, if the robots are in a maze, and they need to decide which landmarks to look at to reduce ambiguity for outlier rejection? I thought a GA would be efficient? Perhaps cons would in a swapc setting it might not be worthwhile...

  • Andrew Davison's profile picture

    @AjdDavison

    Yes that makes sense. Would be fun for sure.

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    Thanks for the encouragement. It’s going to be and I look forward to give an update. Meantime I applied GBP with a factor graph on to a RSSI feeds ( jumpy) fuse with EMF, to localise the source and find directions. Works perfect!

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    flagging here as a thought exercise, I’ve been revisiting the later part of your presentation at NAVER LAB , the “multi vendor interoperability”, I was thinking; let’s assume that robots are learning federated, here the case that the data is Non-IID( none independent and identically distributed) ; Roomba type robot and drone world of view is different , gut is that the standard algorithms would mostly fail, or the SToA highly compute intensive, what if the GA could be much better at finding “good enough” common ground- common translation layer( federated aggregator) , that works for all the robots in the network. Perhaps optimal client selection , etc

    1 like

  • Thiwanka's profile picture

    @ThiwankaCJayasi

    Hi Andrew, managed to get to point, here are my observation with ga, first of all i used euclidian distance, but then again i found your other paper uses the "Mahalanobis distance changes" which is the right thing to do, so I made few changes, what I was trying to solve in this round was to GA-optimized FL with non-IIDd data, solution ga optimized aggregation weights and client selection, result; system adapts to different robot types and improve over rounds, I think real world it would be highly constrained, with varying profiles of terrain, etc. I had a good go thanks to your initial work. Learnt few new concept