Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    SP Mohanty
    @spMohanty
    Hello World :D
    Oleksii
    @Haister
    Hi
    MLerik
    @MLerik
    Hi all. Everything working fine with the repository so far? If you have any questions, don't hesitate to ask. We also encourage you to discuss openly about different topics concerning the challenge.
    Lezwon Castellino
    @Lezwon_twitter
    Hello @MLerik , I've just got started with the challenge. I had on query regarding the rendering. how do I find out which agent is the index in the predictor referring to? I can see the path being decided for an agent handle, but I have no idea which train it's referring to in the rendering
    MLerik
    @MLerik
    Hi @Lezwon_twitter , I don't really understand the question. could you elaborate a bit? The index and the handle are currently the same. thus agent handle 1 will be index one in the predictor as well.
    voanhkha
    @voanhkha
    Hi All, I am happy to participate in this competition. I have been searching all materials on AICrowd and the repo, but can't find where is the training data, and the "submit" button? I can run flatland-rl demos, but what is the exact training scenario, agent's policy format, how to submit...? Could you please help me with this?
    MLerik
    @MLerik
    Hi @voanhkha
    Hope this helps to get started with your submission. If you run into trouble, don't hesitate to ask. :)
    voanhkha
    @voanhkha
    @MLerik Thank you so much for your help!
    nilabha
    @nilabha
    I have put up some code here https://github.com/nilabha/FlatlandCompetition
    This includes actorcritictrainer.py file which implements an actorcritic approach and ESStrategyTraining.py which implements an evolutionary strategy approach
    The results seems to be similar to the Duelling Double DQN approach. I have saved results and pretrained files
    MLerik
    @MLerik

    hi @nilabha : This looks really cool. Since most activity around this challenge is here I suggest you post about this in the forums there.

    It would also be great if you can add some introduction documentation on how to use your repository and comments on your code.

    Have you tried using different observations then the stock observation provided with Flatland? I believe to improve your results using actor critic you need to implement and add new features to your solution. I'm happy to discuss these in the forums with you.

    Best regards,
    Erik

    I am traveling for the next week and a half, so I would pick up working on this after that. In the meantime, any suggestions are most welcome.
    MLerik
    @MLerik
    Really cool. I will have a look at it and am happy to discuss
    I have added another approach which is not model based
    I didn't have success yet with Model based /RL approaches and opted to go for a manual approach which is able to solve some dense networks
    The idea behind using a more manual based approach to get an understanding of which observations are useful.
    nilabha
    @nilabha
    This could be the basis for RL approaches maybe with a bit of reward hacking (Like Adding penalty for deviation from the shortest route for each agent if there are no predicted conflicts). The RL approach shared earlier seems to do poorly with the observations (I tried with some more basic observations also) as I have only one model used for all agents and since the observations of other agents are treated as independent and only the total reward becomes the basis for any gradient updates. I think this can be seen in the degradation in performance as the number of agents increases. A simple shortest path based approach performs better on multiple agents with a sparse railway network.
    @MLerik @spMohanty - Do you think we can have an extension for round 2?
    I would like to see the results of my approaches but It looks difficult that I will have a successful submission by the timeline.
    MLerik
    @MLerik
    @nilabha this sounds very interesting. the round will be closing today but the repository will stay open and we will keep on working to improve both flatland as well as multi-agent approaches for the problem at hand.
    Would you be interested to take an active role in developing flatland further (e.g. Observations and Baseline)?
    nilabha
    @nilabha
    Yes, Sure I would be definitely interested to be in an active role.
    I hope that there would be another competition after this with increasing complexity etc. just like the NIPS Skeleton competitions we have in AICrowd.
    Giulia Cantini
    @giulic3
    Hi, I've a question regarding submissions. Will you keep them open even after the end of Round 2? I am also developing new observations for a DRL approach, I didn't have time to complete within the timeline but I would really like to see how my approach performs against your evaluator :)
    MLerik
    @MLerik
    @giulic3 you will not be able to grade your submissions on the server anymore. You can still use the examples provided on the aicrowd challenge page. Will can still grade some submissions if you need this for publications. please reach out to us directly through email (buttom of page on AIcrowd Challenge Page) if you wish for scoring. We will keep you updated on how submissions can be scored in the future as we have had a few requests about this.
    Giulia Cantini
    @giulic3
    I understand @MLerik , thank you very much for the quick reply!
    Hanchung Lee
    @leehanchung
    Curious, is there a Discord chat for this?
    Florian
    @MasterScrat
    Hey @leehanchung , yes there is! https://discord.gg/hCR3CZG