Skip to content

Inconsistent reward clipping #53

Open
@yunfanjiang

Description

@yunfanjiang

Hi there,

I'm looking at the Optimistic Asymmetric Clipping as mentioned in the IMPALA paper, where a clip function f(r) = 0.3 * min(tanh(r), 0) + 5.0 * max(tanh(r), 0) is used. However, I found it is different from the code implementation (https://github.com/deepmind/scalable_agent/blob/master/experiment.py#L367). In the code above, the clip function is f(r) = 5 * 0.3 * min(tanh(r / 5), 0) + 5 * max(tanh(r / 5), 0), which is obviously not consistent with that in the paper.

I plot these two clip functions. Sadly, it seems that both of them doesn't match Figure D.1. in the paper.
image

Could you please tell which clip function was used in the experiments? And also, any explanations about the discrepancy will be helpful!

Thanks for your reading and I look forward to hearing from you.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions