Skip to content

The use of quality value and weighted reward. #28

@OneCoin123

Description

@OneCoin123

Thank you for your elegent works!

According to my understanding, you proposed a form of weighted reward w_{s_k} for thought expanding actions and quality value v_k for node state, which are used to guide the searching in MCTS.

Referring to your paper, w_{s_k} is computed from the distance, parent node value, and the reward estimation for the generated thought, and v_k is computed from the parent node value and w_{s_k}.

However, I feel confused since the MCTS code in this repo seems to use the LLM-generated reward directly as the child node's value, skipping your design of w and v in the paper.

Did I miss something, or misunderstand your pipeline? Looking forward to your kind response.

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