search

beanie00 / Decision-ConvFormer

[ICLR 2024 Spotlight] Code for the paper "Decision ConvFormer: Local Filtering in MetaFormer is Sufficient for Decision Making"
https://www.beanie00.com/publications/dc
MIT License
8 stars 1 forks source link

Atari score #5

Open WangJinCheng1998 opened 1 month ago

WangJinCheng1998 commented 1 month ago

Hi! I am running the DC on the Atari Breakout game. Unfortunately, I got an average of 320 normalized scores on 5 seeds by using DChytbrid. Would you mind providing the human score and the random score that was used to normalize the results? Thank you so much.

beanie00 commented 1 month ago

Thank you for the question. We obtained the Atari game random and gamer scores from 'Mastering Atari with Discrete World Models'[1]. The scores can be found in the table on page 26. For Breakout, we used a random score of 2 and a gamer score of 30.

Please note that although we mentioned in the paper that we normalized the scores according to 'Mastering Atari Games with Limited Data,'[2] the correct source is 'Mastering Atari with Discrete World Models.' We will update the arXiv version accordingly. The random and gamer scores from both papers are nearly identical but differ slightly in decimal points.

[1] Hafner, Danijar, et al. 'Mastering Atari with Discrete World Models.' ICLR 2021. [2] Ye, Weirui, et al. 'Mastering Atari Games with Limited Data.' NeurIPS 2021.