thu-ml/tianshou

Extend and fix reward/return normalizations

Open

#927 opened on 2023年8月29日

GitHub で見る
 (0 comments) (0 reactions) (0 assignees)Python (7,121 stars) (1,072 forks)batch import
enhancementgood first issue

説明

There are some ways in which reward/return/value normalization could be improved. But one most drastic thing is the following:

Currently PGPolicy instantiates self.ret_rms = RunningMeanStd(), and RunningMeanStd has a default value of clip_max=10. This cannot be adjusted by users! (except through monkey-patching, ofc)

This might work well for some standard envs, but the clipping value is arbitrary and making it non-configurable is a major hinderance for users, who are most probably not aware of this.

Generally, how to best normalize stuff in RL is an active discussion and normalization can play an important role in performance. I believe tianshou should be extended to accomodate various normalization strategies

コントリビューターガイド

技術スタック
pythonpytorch
領域
machine learningai
Issue 種別
feature
難度新しい貢献者にとっての実装難度です。1 は非常に小さな変更、5 は専門的な作業です。
3
推定時間経験ある貢献者が調査、実装、テスト、pull request 準備にかけるおおよその時間範囲です。
1-2 days
活動状況Issue が今どれだけ取り組みやすいかを示します。新しい、活発、古い、ブロック中、またはメンテナー入力待ちなどです。
fresh
明確さ期待される変更、受け入れ条件、次の手順がどれだけ明確かを示します。
mostly clear
前提条件
understanding of RL reward normalizationfamiliarity with tianshou codebase
初心者向け度初回貢献者にどれだけ取り組みやすいかを 1-100 で推定したスコアです。
40
調査方針
Investigate the current normalization implementation in PGPolicy, specifically the `RunningMeanStd` class with default `clip max=10` in `tianshou/policy/pg.py`. The goal is to make `clip max` configurable and possibly add support for alternative normalization strategies. Review the entire codebase for similar hardcoded normalization values and consider creating a unified configuration interface.