.png){kind=icon}
.png){kind=icon}
Return
Dynamic subtask representation and assignment in cooperative multi-agent tasks
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.129535.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In cooperative multi-agent reinforcement learning, achieving scalability is a critical objective, often pursued through parameter sharing. However, parameter sharing often leads to homogeneous agent behaviors, thereby reducing strategic diversity. Task decomposition emerges as a viable solution to achieve both scalability and diversity in multi-agent systems, but existing methods struggle with intricate environments due to reliance on prior knowledge or ignoring agent-specific characteristics. To bridge this gap, we introduce a novel framework for learning Dynamic Subtask Representation and Assignment (DSRA) in cooperative multi-agent tasks. We first leverage a variational autoencoder with a multivariate Gaussian distribution, enabling agents to generate dynamic and distinct subtask representations. After that, we incorporate an ability encoder to optimize the subtask assignment process in alignment with each agent's abilities. To encourage diversity while sharing experiences, we further empower agents to develop individual policies for each subtask, promoting specialization and encouraging knowledge sharing, especially among agents engaged in similar tasks. Empirical results on Level-based Foraging, Predator-Prey, and Starcraft II micromanagement challenges demonstrate that DSRA significantly improves learning performance compared to existing MARL methods.
Keywords:
Multi-agent system
Reinforcement learning
Task decomposition