Construct A Unique Agent Network For Cooperative Multi -agent Reinforcement Learning
Long Liang, Haolin Wu, Hui Li · 2022
Cooperative multi-agent reinforcement learning based on centralized training decentralized execution (CTDE) has shown outstanding performance in handling cooperative multi-agent tasks. However, the agent network in CTDE-based algorithms is a shared agent network, which cannot help each agent make an accurate decentralized policy. Simultaneously, constructing a network for each agent would increase the complexity of the algorithms and make it difficult to train these agent networks in an end-to-end fashion. To solve such problems, we propose a unique method for constructing agent networks, which uses a hypernetwork to dynamically generate weights for the agent network according to the current inputs. So that the weight parameters of the agent networks can float with the inputs and adapt to the input sequence. In this way, each agent network can accurately make a decentralized policy in the current agent state. Our experimental results show that the proposed method can improve the convergence efficiency and the policy performance of cooperative multi-agent reinforcement learning algorithms.