Search Unity

Adding new training to finished brains and will they keep learning in playable game?

Discussion in 'ML-Agents' started by Bjuglez, Mar 27, 2020.

  1. Bjuglez

    Bjuglez

    Joined:
    Apr 25, 2014
    Posts:
    5
    I am interested in 2 things on how this works.

    If I train a brain to do a simple task like pick up object or something and finish and then decide to add actions to that same brain and train it again does it still know what it learned since last time? Can i for example teach it to pick up, ok finish, then add eat and then add mining etc to improve it. Or better yet teach it each of those individually in environments designed to those specific tasks and then merge them all together?

    Also if I make a game with something like this and people play my game, will the final trained brain keep on learning when the game is being played or will the brain just keep its functionality that it got at the end of the training I did for it while in the unity editor?
     
    Bjuglez, Mar 27, 2020
    #1

  2. christophergoy

    christophergoy

    Unity Technologies

    Joined:
    Sep 16, 2015
    Posts:
    735
    The short answer to your first question is no. You would need to have all of your observations and actions setup initially and do something like curriculum learning in order to ramp up learning over time.

    The answer to your second question is also no. A trained and deployed brain does not learn. It is essentially "frozen" once you are done training it.
     
    christophergoy, Mar 30, 2020
    #2

  3. LexVolkov

    LexVolkov

    Joined:
    Sep 14, 2014
    Posts:
    62
    in the future, it will be impossible to "finish off" the brain?
     
    LexVolkov, Mar 30, 2020
    #3

  4. LexVolkov

    LexVolkov

    Joined:
    Sep 14, 2014
    Posts:
    62


    Does this option not help?
     
    LexVolkov, Mar 31, 2020
    #4

  5. christophergoy

    christophergoy

    Unity Technologies

    Joined:
    Sep 16, 2015
    Posts:
    735
    You can resume a training session, which is different than a brain learning that is deployed in a game. There is currently no support for 'online' learning in ml-agents.
     
    christophergoy, Apr 1, 2020
    #5