起始问题
提问原文
As I understand, in some approaches, agents which are RL-trained using natural language to communicate are required to be trained together to have the same protocol for the communication. Is there any work that does not need RL-agents being trained together?