Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames

Zhikai Zhou; Hongbin Ma; Ying Jin; Yehao Fang; Haipeng Wang; Rufei Zhang

doi:10.1007/978-981-95-6733-1_13

Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames

Zhikai Zhou^*
, Hongbin Ma
, Ying Jin
, Yehao Fang
, Haipeng Wang
, Rufei Zhang^*

^*Corresponding author for this work

School of Automation

Beijing Institute of Technology
National Key Lab of Autonomous Intelligent Unmanned Systems
Beijing Institute of Control and Electronics Technology

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

To address the critical need for robust and accurate decision-making in complex, real-time environments, we propose a lightweight decision framework that dynamically selects the best candidate from an algorithm pool comprising Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM) and Deep Q-Network (DQN). A trust-based selector first picks the most reliable model in real time; a further model-fusion step then re-weights the outputs of all models to push accuracy beyond the best single model. Experiments on the “Miao-Suan” tactical wargame show that the framework raises average accuracy from 95.47% (best single model) to 99.82% while cutting variance by two orders of magnitude, without extra training cost.

Original language	English
Title of host publication	Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings
Editors	Hongbin Ma, Bin Xin, Qing Wang, Jinhua She
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	157-169
Number of pages	13
ISBN (Print)	9789819567324
DOIs	https://doi.org/10.1007/978-981-95-6733-1_13
Publication status	Published - 2026
Event	9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025 - Zhuhai, China Duration: 31 Oct 2025 → 4 Nov 2025

Publication series

Name	Communications in Computer and Information Science
Volume	2781 CCIS
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025
Country/Territory	China
City	Zhuhai
Period	31/10/25 → 4/11/25

Keywords

Algorithm pool
Model selection
Reinforcement learning

Access to Document

10.1007/978-981-95-6733-1_13

Cite this

Zhou, Z., Ma, H., Jin, Y., Fang, Y., Wang, H., & Zhang, R. (2026). Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames. In H. Ma, B. Xin, Q. Wang, & J. She (Eds.), Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings (pp. 157-169). (Communications in Computer and Information Science; Vol. 2781 CCIS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-95-6733-1_13

Zhou, Zhikai ; Ma, Hongbin ; Jin, Ying et al. / Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames. Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings. editor / Hongbin Ma ; Bin Xin ; Qing Wang ; Jinhua She. Springer Science and Business Media Deutschland GmbH, 2026. pp. 157-169 (Communications in Computer and Information Science).

@inproceedings{8557fff4331649b78859e3e723168fee,

title = "Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames",

abstract = "To address the critical need for robust and accurate decision-making in complex, real-time environments, we propose a lightweight decision framework that dynamically selects the best candidate from an algorithm pool comprising Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM) and Deep Q-Network (DQN). A trust-based selector first picks the most reliable model in real time; a further model-fusion step then re-weights the outputs of all models to push accuracy beyond the best single model. Experiments on the “Miao-Suan” tactical wargame show that the framework raises average accuracy from 95.47\% (best single model) to 99.82\% while cutting variance by two orders of magnitude, without extra training cost.",

keywords = "Algorithm pool, Model selection, Reinforcement learning",

author = "Zhikai Zhou and Hongbin Ma and Ying Jin and Yehao Fang and Haipeng Wang and Rufei Zhang",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.; 9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025 ; Conference date: 31-10-2025 Through 04-11-2025",

year = "2026",

doi = "10.1007/978-981-95-6733-1\_13",

language = "English",

isbn = "9789819567324",

series = "Communications in Computer and Information Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "157--169",

editor = "Hongbin Ma and Bin Xin and Qing Wang and Jinhua She",

booktitle = "Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings",

address = "Germany",

}

Zhou, Z, Ma, H , Jin, Y, Fang, Y, Wang, H & Zhang, R 2026, Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames. in H Ma, B Xin, Q Wang & J She (eds), Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings. Communications in Computer and Information Science, vol. 2781 CCIS, Springer Science and Business Media Deutschland GmbH, pp. 157-169, 9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025, Zhuhai, China, 31/10/25. https://doi.org/10.1007/978-981-95-6733-1_13

Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames. / Zhou, Zhikai; Ma, Hongbin ; Jin, Ying et al.
Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings. ed. / Hongbin Ma; Bin Xin; Qing Wang; Jinhua She. Springer Science and Business Media Deutschland GmbH, 2026. p. 157-169 (Communications in Computer and Information Science; Vol. 2781 CCIS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames

AU - Zhou, Zhikai

AU - Ma, Hongbin

AU - Jin, Ying

AU - Fang, Yehao

AU - Wang, Haipeng

AU - Zhang, Rufei

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

PY - 2026

Y1 - 2026

N2 - To address the critical need for robust and accurate decision-making in complex, real-time environments, we propose a lightweight decision framework that dynamically selects the best candidate from an algorithm pool comprising Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM) and Deep Q-Network (DQN). A trust-based selector first picks the most reliable model in real time; a further model-fusion step then re-weights the outputs of all models to push accuracy beyond the best single model. Experiments on the “Miao-Suan” tactical wargame show that the framework raises average accuracy from 95.47% (best single model) to 99.82% while cutting variance by two orders of magnitude, without extra training cost.

AB - To address the critical need for robust and accurate decision-making in complex, real-time environments, we propose a lightweight decision framework that dynamically selects the best candidate from an algorithm pool comprising Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM) and Deep Q-Network (DQN). A trust-based selector first picks the most reliable model in real time; a further model-fusion step then re-weights the outputs of all models to push accuracy beyond the best single model. Experiments on the “Miao-Suan” tactical wargame show that the framework raises average accuracy from 95.47% (best single model) to 99.82% while cutting variance by two orders of magnitude, without extra training cost.

KW - Algorithm pool

KW - Model selection

KW - Reinforcement learning

UR - https://www.scopus.com/pages/publications/105038480395

U2 - 10.1007/978-981-95-6733-1_13

DO - 10.1007/978-981-95-6733-1_13

M3 - Conference contribution

AN - SCOPUS:105038480395

SN - 9789819567324

T3 - Communications in Computer and Information Science

SP - 157

EP - 169

BT - Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings

A2 - Ma, Hongbin

A2 - Xin, Bin

A2 - Wang, Qing

A2 - She, Jinhua

PB - Springer Science and Business Media Deutschland GmbH

T2 - 9th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2025

Y2 - 31 October 2025 through 4 November 2025

ER -

Zhou Z, Ma H , Jin Y, Fang Y, Wang H, Zhang R. Online Model-Pool Selection and Fusion for Adaptive MARL in Wargames. In Ma H, Xin B, Wang Q, She J, editors, Advanced Computational Intelligence and Intelligent Informatics - 9th International Workshop, IWACIII 2025, Proceedings. Springer Science and Business Media Deutschland GmbH. 2026. p. 157-169. (Communications in Computer and Information Science). doi: 10.1007/978-981-95-6733-1_13