Research on Generalization of Multi-Agent Based on CMDP

• 1、School of Artificial Intelligence, Beijing University Of Posts and Telecommunications, BeiJing 010000

Abstract：Aiming at the problem of multi-agent task generalization, this paper proposes a multi-agent reinforcement learning algorithm based on context Markov decision. In this algorithm, a context information module is designed to realize the representation of context by using the attention mechanism. Context information contains additional task-related information, which can be used to infer the relationship between different tasks under the multi-task setting. In this way, the experience of historical tasks can be saved and utilized in new tasks, so as to realize the generalization between tasks. In addition, the action subspace is designed based on the context information, which solves the space explosion problem and accelerates the efficient training of the agent. Compared with the existing algorithm in the multi-agent particle environment, the experimental results show that the proposed method is more effective in other tasks.

Keywords： Reinforcement learning Multi-agent CMDP Generalization