Multi-Agent Reinforcement Learning for Large Model-Aided Mobile Applications

Yiwen Zhan

Keywords: Multi-agent reinforcement learning, large model, communications, mobile applications, wireless networks

TL;DR: We propose an attention-based multi-agent reinforcement learning scheme that optimizes communication and experience sharing to improve the quality of service for mobile applications in wireless networks.

Abstract: Multi-agent reinforcement learning utilizes the observations and learning experiences shared among the agents to accelerate optimization efficiency under partial observations, but the performance degrades due to the excessive communications and large volumes of irrelevant experiences. In this paper, we propose a multi-agent reinforcement learning scheme to optimize communication decision, the cooperative agent selection, and the required number of sharing experiences based on the task features and previous contributions of neighbors extracted by large model to improve the quality of service for mobile applications in wireless networks. The experiences with high temporal difference error are shared and utilized based on the attention mechanism, which extracts the correlation with the reachable neighbors set to improve optimization efficiency. As a case study, the proposed communication scheme is implemented in the vision transformer-aided collaborative perception system based on connected autonomous vehicles and roadside unit to support 3D object detection, and the performance gain is verified via simulation results.

Submission Number: 4