Search Unity

How do I use Curriculum Learning with Group Collaborative rewarded Agents ?

Discussion in 'ML-Agents' started by JulesVerny, Feb 8, 2022.

  1. JulesVerny

    JulesVerny

    Joined:
    Dec 21, 2015
    Posts:
    47
    The Unity documentation on the Configuration file for Curriculum learning is rather weak, but I have found examples I can use for single Agent training, to advance the environment parameters based upon reward measure. (c.f. measure: reward)

    However now I have a multi agent, environment where the effective reward signal is now a Group Cumulative Reward signal. However it is not obvious how I use this signal to advance the curriculum environment parameters, as my single Cumulative Reward signal remains flat zero, whilst my Group Cumulative Reward is actually advancing.

    How do I distinguish to use the Group Cumulative Reward as the measure for level progress the configuration file ?
    "measure: reward" does not seem to work work, for the Group Culmulative Reward

    It is not obvious in the documentation what the correct measure name would be. None of the poca collaborative examples (which tend to use self play and not any Curriculum learning)

    Cheers for any help
     
    Last edited: Feb 8, 2022
  2. almostgiants

    almostgiants

    Joined:
    Nov 10, 2021
    Posts:
    5
    Just curious if you ever found the answer? I had the same question. Thanks!!