|
|
|
| Research on Multi-Agent Training for Maritime Joint Air Defence Based on a Multi-Algorithm Framework with Adaptive Hierarchical Sharing |
| YE Qichang1， WAN Shizheng2， LI Yueshu1， CHEN Zhumei1， LIU Shanglin1 |
| 1. School of Information and Communication Engineering， University of Electronic Science
and Technology of China， Chengdu 611731， Sichuan， China；
2. Shanghai Electro-Mechanical Engineering Institute， Shanghai 201109， China |
|
|
|
|
Abstract Multi-agent training is widely applied in modern maritime joint air defence operations. When two opposing parties independently choose different learning algorithms, it is often necessary to ensure implementation consistency and to manage resource allocation. To achieve this, a cross-agent shared fully connected layer is typically employed as a common representational foundation. However, a static choice between “fully shared” and “fully private” structures fails to balance policy consistency with individual differentiation. At the same time, discrepancies in behaviour distribution and gradient statistics across algorithms may amplify negative transfer and training variance. To overcome the above challenges, this paper introduced adaptive layer sharing (ALS) across heterogeneous algorithms, enabling learnable gating mechanisms to dynamically weight between shared and private branches at every layer. In a small-scale, single-machine experimental setup, standardized and reproducible protocols were established to record and report game outcomes and compliance. When ALS was activated, the learned gating distributions and thresholded topologies would be extracted, creating an implementable and observable engineering baseline that provides a clear structural and metric foundation for future large-scale and multi-task evaluations.
|
|
Received: 30 October 2025
Published: 13 January 2026
|
|
|
|
|
|
| [1] |
ZHANG Wanying, SIMA Ke, ZHANG Yuhe, MENG Jian, YANG Zhen, ZHOU Deyun. The Guidance and Control Method of Multi-Missile Cooperative Encirclement of Maneuvering Targets Based on Proximal Policy Optimization[J]. Air & Space Defense, 2025, 8(4): 94-103. |
| [2] |
QUAN Jiale, MA Xianlong, SHEN Yuheng. Multi-agent Formation Method Based on Dynamic Optimization of Proximal Policies[J]. Air & Space Defense, 2024, 7(2): 52-62. |
| [3] |
LI Mengxuan, GUO Jianguo, XU Xinpeng, SHEN Yuheng. Guidance Law Based on Proximal Policy Optimization[J]. Air & Space Defense, 2023, 6(4): 51-57. |
|
|
|
|
2025, Vol. 8 
