Dynamic role discovery and assignment in multi-agent task decomposition

Abstract

Effective multi-agent teamwork can be facilitated by using personas to decompose goals into lower-level team subtasks through a shared understanding of multi-agent tasks. However, traditional methods for role discovery and assignment are not scalable and fail to adapt to dynamic changes in the environment. To solve this problem, we propose a new framework for learning dynamic role discovery and assignment. We first introduce an action encoder to construct a vector representation for each action based on its characteristics, and define and classify roles from a more comprehensive perspective based on both action differences and action contributions. To rationally assign roles to agents, we propose a representation-based role selection policy based on consideration of role differences and reward horizons, which enables agents to play roles dynamically by dynamically assigning agents with similar abilities to play the same role. Agents playing the same role share their learning of the role, and different roles correspond to different action spaces. We also introduce regularizers to increase the differences between roles and stabilize training by preventing agents from changing roles frequently. Role selection and role policy integrate action representations and role differences in a restricted action space, improving learning efficiency. We conducted experiments in the SMAC benchmark and showed that our method enables effective role discovery and assignment, outperforming the baseline on four of the six scenarios, with an average improvement in win rate of 20%, and is effective in hard and super hard maps. We also conduct ablation experiments to demonstrate the importance of each component in our approach.