TonghanWang / RODE

Codes accompanying the paper "RODE: Learning Roles to Decompose Multi-Agent Tasks (ICLR 2021, https://arxiv.org/abs/2010.01523). RODE is a scalable role-based multi-agent learning method which effectively discovers roles based on joint action space decomposition according to action effects, establishing a new state of the art on the StarCraft multi-agent benchmark.
Apache License 2.0
69 stars 20 forks source link

您对环境(starcraft II.py)修改的思路 #10

Closed zyh1999 closed 2 years ago

zyh1999 commented 2 years ago

您好,我认真的阅读了您发表的rode这篇文章,并尝试复现了您的代码,在复现的过程中发现似乎在您的(starcraft II.py)文件中的init_units函数里加入的对敌人观察的排序,您还特别加上了这段是您写的的注释。当我在原始的https://github.com/oxwhirl/smac/tree/master/smac/env/starcraft2 环境里加上这个初始化排序,发现对效果会有一定的影响。想请问一下您对此处修改的理由?

TonghanWang commented 2 years ago

您好,这是因为starcraft这个环境,action和敌人有关(一个attack action针对一个特定敌人),而rode在学action representation的时候,一个action id必须对应固定的一个action。

这个环境每次reset时,同一个action id可能会攻击不同的敌人,导致action语意变化了,这就没法学好action representation,所以必须排序。

On Sun, May 8, 2022 at 9:28 PM zyh1999 @.***> wrote:

您好,我认真的阅读了您发表的rode这篇文章,并尝试复现了您的代码,在复现的过程中发现似乎在您的(starcraft II.py)文件中的init_units函数里加入的对敌人观察的排序,您还特别加上了这段是您写的的注释。当我在原始的 https://github.com/oxwhirl/smac/tree/master/smac/env/starcraft2 环境里加上这个初始化排序,发现对效果会有一定的影响。想请问一下您对此处修改的理由?

— Reply to this email directly, view it on GitHub https://github.com/TonghanWang/RODE/issues/10, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIX4MOX4INHOFXKDGTZXZNTVI66O5ANCNFSM5VL57RSA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

zyh1999 commented 2 years ago

谢谢您的回复,想了一下确实是这么回事,难怪我之前尝试复现的时候只加入动作表示不加入角色相关的组件并没有得到比baseline更好的效果(没排序)。另外恳请您也能看一下邮箱里我更详细的问题描述~

TonghanWang commented 2 years ago

只加action representation确实应该不咋行,因为action representation是有语义的,直接用语义空间和策略空间没法对齐.

您的邮件我已仔细阅读并回复😄

On Sun, May 8, 2022 at 9:55 PM zyh1999 @.***> wrote:

谢谢您的回复,想了一下确实是这么回事,难怪我之前尝试复现的时候只加入动作表示不加入角色相关的组件并没有得到比baseline更好的效果(没排序)。另外恳请您也能看一下邮箱里我更详细的问题描述~

— Reply to this email directly, view it on GitHub https://github.com/TonghanWang/RODE/issues/10#issuecomment-1120423375, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIX4MOXNWYPCI337P4LGU6DVI7BTTANCNFSM5VL57RSA . You are receiving this because you commented.Message ID: @.***>