Skip to content

RL training does not converge #30

Description

@cjywokok3550

Hi, thanks for your great work!

I am trying to reproduce the training results using the default settings provided in the repository. However, I encountered an issue related to training stability.

According to the paper, the training reward shows a steady upward trend throughout the process. However, in my experiments, the reward behaves differently: The reward increases at the beginning. Around ~500 episodes (after the second stage of curriculum learning starts), the reward drops significantly. After that, the training becomes unstable and fails to converge

I would like to ask: Is this drop in reward during curriculum transition expected? Are there any sensitive hyperparameters that could affect stability? Could this be related to environment or dependency versions?

Additionally, I did not modify the original code or hyperparameters.

I would really appreciate any suggestions on how to stabilize the training.
Thank you!

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