Search Unity

MA-POCA PushBlockCollab

Discussion in 'ML-Agents' started by JHeloise, May 20, 2022.

  1. JHeloise

    JHeloise

    Joined:
    Nov 19, 2020
    Posts:
    2
    Help wanted!!!
    I tried to train PushBlockCollab project using the yaml file in directory .\config\poca\PushBlockCollab.yaml, but the result was not well. The group cumulative reward curve is shown below. The reward is about 1.4 even after 12 million training steps. I did not change anything in the project. I can not understand why my training failed.
     

    Attached Files:

    JHeloise, May 20, 2022
    #1

  2. iffalseelsetrue

    iffalseelsetrue

    Joined:
    May 3, 2018
    Posts:
    11
    yeah I also noticed the very low value of the Group Cumulative Reward, I have a test project setup using POCA and the Group Cumulative Reward was hitting 900...

    would like to add a question here
    what's the normal range of the Group Cumulative Reward?
     
    iffalseelsetrue, May 31, 2022
    #2

  3. JHeloise

    JHeloise

    Joined:
    Nov 19, 2020
    Posts:
    2
    I think the value of Group Cumulative Reward depends on the the value of rewards you set in the scripts.
    My problem is why my training result of Group Cumulative Reward is far below the what is described in the documentation.
    https://github.com/Unity-Technologies/ml-agents/blob/main/docs/Learning-Environment-Examples.md
     

    Attached Files:

    JHeloise, Jun 7, 2022
    #3

  4. ice_creamer

    ice_creamer

    Joined:
    Jul 28, 2022
    Posts:
    34
    Hi, the paper ma-poca last part can see about trainer.yaml. Do you know how to write RSA setting in yaml?
     
    ice_creamer, Jul 12, 2023
    #4