Search Unity

Resolved Train a ML-agent with # curiosity module

Discussion in 'ML-Agents' started by XXXXchris, Apr 6, 2021.

  1. XXXXchris

    XXXXchris

    Joined:
    Jan 24, 2021
    Posts:
    10
    Hi, I used ML-agent train agent with curiosity, so How can I record the agent’s intrinsic and extrinsic rewards for every observation and timestep?
     
    XXXXchris, Apr 6, 2021
    #1

  2. XXXXchris

    XXXXchris

    Joined:
    Jan 24, 2021
    Posts:
    10
    or how do I record the rewards for every observation and timestep?
     
    XXXXchris, Apr 9, 2021
    #2

  3. andrewcoh_unity

    andrewcoh_unity

    Unity Technologies

    Joined:
    Sep 5, 2019
    Posts:
    162
    The cumulative reward per timestep for all reward signals is reported in tensorboard. Are you asking for a different statistic than this?
     
    andrewcoh_unity, Apr 13, 2021
    #3

  4. XXXXchris

    XXXXchris

    Joined:
    Jan 24, 2021
    Posts:
    10
    Yes, when I use curiosity configuration, how can I distinguish between intrinsic and extrinsic rewards? so In Ml-agent, the only statistic I can know is the cumulative reward per timestep for all reward signals?
     
    XXXXchris, Apr 13, 2021
    #4

  5. andrewcoh_unity

    andrewcoh_unity

    Unity Technologies

    Joined:
    Sep 5, 2019
    Posts:
    162
    If you open the Policy tab in tensorboard, you will see Curiosity Reward and Extrinsic Reward as separate plots.
     
    andrewcoh_unity, Apr 13, 2021
    #5

  6. XXXXchris

    XXXXchris

    Joined:
    Jan 24, 2021
    Posts:
    10
    got it, thanks.
     
    XXXXchris, Apr 14, 2021
    #6