Skip to content

Normal on-policy converges to the best reward but ACER jumps around a lower value #5

@kadhirumasankar

Description

@kadhirumasankar

Hi Didier, I started two simultaneous runs: one that used the replay buffer and one that did not. I noticed that initial learning moved a lot quicker to higher scores when using ACER, but it eventually started flattening and slowing down, and was overtaken by the normal on-policy run. I just wanted to ask you if you knew of any hyperparameters or tweaks to the code from your experience that might change this behavior and make ACER converge quicker than normal on-policy.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions