Learning to Communicate Strategically for Efficient Collective Intelligence
Xinghai Wei, Tingting Yuan, Jie Yuan, Xiaoming Fu · 2024
Learning to communicate (L2C) involves learning how, when, and with whom to communicate to enhance cooperation among agents under limited bandwidth. However, introducing L2C impedes the original learning tasks, resulting in slower learning and poor convergence, indicative of multitask interference. To address these issues, we propose a strategic integration of L2C with the original learning tasks. This approach enables progressive learning by segmenting the training process into discrete or continuous courses. These courses are structured to become progressively more challenging by gradually increasing the emphasis on rewards linked to communication.