Search Unity

  1. Megacity Metro Demo now available. Download now.
    Dismiss Notice
  2. Unity support for visionOS is now available. Learn more in our blog post.
    Dismiss Notice

Training versus a pretrained Brain

Discussion in 'ML-Agents' started by Sephraih, Apr 13, 2020.

  1. Sephraih

    Sephraih

    Joined:
    Jun 15, 2019
    Posts:
    11
    Of course, self-play already covers this during the training, however, I would love to be able to reuse a trained brain to pit a variety of new setups against each other - creating a basic ai to compete against for updated or specialized agents (that have different abilities for example, or learn with different rewards, etc.)

    So far, I have tested the trained models by adding them to the Behavior Parameters under the model - adding them there and starting a new training will however train a new brain as far as i know.
    Having a different gameobject with the same setup, would still be subject to that training, hence, seeing I wasn't able to find a parameter to be given to the --train command that would allow to only train a certain brain / behavior depending on its name or so, I wondered if there is any way to achieve this at all - is there for example a component that only plays the brain but doesn't train them?

    Thanks a lot for your time and replies! :)
     
  2. andrewcoh_unity

    andrewcoh_unity

    Unity Technologies

    Joined:
    Sep 5, 2019
    Posts:
    162
    Can you use the 'inference only' setting of the behavior type in the behavior parameters script to achieve this?
     
  3. Sephraih

    Sephraih

    Joined:
    Jun 15, 2019
    Posts:
    11
    Ah, thanks, seems like i misunderstood the inference option to train if connected to a trainer.
    - oddly, if i keep one of the self-play agents set to inference only and another to train anew (default), the model of the training agent seems to update to the interference agent brain's behavior quite rapidly, which i deemed it not working earlier - I'll get to the bottom of it, i guess :)

    thanks again, greetings,
    seph