基于模型融合和生成网络的有效阵位智能决策方法

doi:10.16182/j.issn1004731x.joss.22-1265

摘要/Abstract

摘要：

军事智能技术是当前最具活力的前沿领域和未来无人装备发展的必然趋势。针对无人平台在复杂环境下自主决策可靠性和实时性的双重需求和现有基于规则推演的作战仿真技术在动态性和灵活性方面的不足，采用原理分析与实验验证的研究方法，在某型无人平台射击实验数据集的基础上，围绕攻击决策的有效阵位识别环节，将其转换为机器学习领域类别不平衡的二分类问题，综合采用相关性分析、特征工程、模型融合技术构建高实时性和灵活性的有效阵位智能决策模型，并提出基于ICGAN-Stacking不平衡分类架构对少数类样本进行定向扩充，实现数据增强和模型性能提升。实验结果表明：所提方法召回率提升了4.1%、精确度提升了0.4%、F1值提升了1.5%、AUC值达到90.9%，能够满足无人平台执行作战任务实时性和可靠性需求。

关键词: 军事智能, 无人平台, 模型融合, 生成对抗网络, 不平衡分类

Abstract:

Military intelligence technology is currently the most dynamic frontier and the inevitable trend for the development of unmanned equipment in the future. Aiming at the dual requirements of reliability and real-time performance of unmanned platform autonomous decision-making in complex environments and the shortcomings of existing combat simulation technology based on rule reasoning in terms of dynamics and flexibility, a research method of principle analysis and experimental verification is adopted. Based on the shooting experiment dataset of an unmanned platform, the effective position recognition link of attack decision-making is transformed into a binary classification problem with imbalanced categories in the field of machine learning. The effective position intelligent decision-making model with high real-time performance and flexibility is constructed by using correlation analysis, feature engineering, and model fusion technology. Based on the imbalanced classification architecture of ICGAN-Stacking, directional expansion of minority class samples is proposed to achieve data enhancement and model performance improvement. The experimental results show that the recall rate of the proposed method has increased by 4.1%, the accuracy by 0.4%, and the F1 value by 1.5%, and the AUC value reaches 90.9%, which can meet the real-time performance and reliability requirements of the unmanned platform in performing combat tasks.

Key words: military intelligence, unmanned platform, model fusion, generative adversarial network, imbalance classification

中图分类号:

TP181

郭力强,马亮,张会等 . 基于模型融合和生成网络的有效阵位智能决策方法[J]. 系统仿真学报, 2024, 36(7): 1573-1585.

Guo Liqiang,Ma Liang,Zhang Hui,et al . Effective Position Intelligent Decision Method Based on Model Fusion and Generative Network[J]. Journal of System Simulation, 2024, 36(7): 1573-1585.

图/表 17

表1

图1

图2

表2

图3

图4

图5

图6

表3

表4

图7

表5

图8

图9

图10

图11

表6

参考文献 37

1	司广宇, 苗艳, 李关防. 水下立体攻防体系构建技术[J]. 指挥控制与仿真, 2018, 40(1): 1-8.
	Si Guangyu, Miao Yan, Li Guanfang. Underwater Tridimensional Attack-defense System Technology[J]. Command Control and Simulation, 2018, 40(1): 1-8.
2	何玉庆, 秦天一, 王楠. 跨域协同:无人系统技术发展和应用新趋势[J]. 无人系统技术, 2021, 4(4): 1-13.
	He Yuqing, Qin Tianyi, Wang Nan. Cross-domain Collaboration: New Trends in the Development and Application of Unmanned Systems Technology[J]. Unmanned Systems Technology, 2021, 4(4): 1-13.
3	王雅琳, 杨依然, 王彤, 等. 2019年无人系统领域发展综述[J]. 无人系统技术, 2019, 2(6): 53-57.
	Wang Yalin, Yang Yiran, Wang Tong, et al. Summary of the Development of Unmanned Systems in 2019[J]. Unmanned Systems Technology, 2019, 2(6): 53-57.
4	李磊, 王彤, 蒋琪. 从美军2042年无人系统路线图看无人系统关键技术发展动向[J]. 无人系统技术, 2018, 1(4): 79-84.
	Li Lei, Wang Tong, Jiang Qi. Key Technology Develop Trends of Unmanned Systems Viewed from Unmanned Systems Integrated Roadmap 2017-2042[J]. Unmanned Systems Technology, 2018, 1(4): 79-84.
5	周光霞, 周方. 美军人工智能空战系统阿尔法初探[C]//第六届中国指挥控制大会论文集(上册). 北京: 电子工业出版社, 2018: 66-70.
6	王建丽, 张渭育. 统计学[M]. 北京: 清华大学出版社, 2010: 215-220.
7	周志华. 机器学习[M]. 北京: 清华大学出版社, 2016.
	Zhou Zhihua. Machine Learning[M]. Beijing: Tsinghua University Press, 2016.
8	Quinlan J R. Induction of Decision Trees[J]. Machine Learning, 1986, 1(1): 81-106.
9	Olson R S, Moore J H. Identifying and Harnessing the Building Blocks of Machine Learning Pipelines for Sensible Initialization of a Data Science Automation Tool[M]//Riolo R, Worzel B, Goldman B, et al. Genetic Programming Theory and Practice XIV. Cham: Springer International Publishing, 2018: 211-223.
10	冯国双. 白话统计[M]. 北京: 电子工业出版社, 2018.
11	Peng Hanchuan, Long Fuhui, Ding C. Feature Selection Based on Mutual Information: Criteria of Max-dependency, Max-relevance, and Min-redundancy[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(8): 1226-1238.
12	Cortes C, Vapnik V. Support-vector Networks[J]. Machine Learning, 1995, 20(3): 273-297.
13	Narasimhamurthy A M. A Framework for the Analysis of Majority Voting[C]//Image Analysis. Berlin: Springer Berlin Heidelberg, 2003: 268-274.
14	Wolpert D H. Stacked Generalization[J]. Neural Networks, 1992, 5(2): 241-259.
15	Kuo C C J. Understanding Convolutional Neural Networks with a Mathematical Model[J]. Journal of Visual Communication and Image Representation, 2016, 41: 406-413.
16	夏恒, 汤健, 乔俊飞. 深度森林研究综述[J]. 北京工业大学学报, 2022, 48(2): 182-196.
	Xia Heng, Tang Jian, Qiao Junfei. Review of Deep Forest[J]. Journal of Beijing University of Technology, 2022, 48(2): 182-196.
17	Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative Adversarial Networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
18	Kingma Diederik P, Welling Max. Auto-encoding Variational Bayes[C]//ICLR 2014. New York, USA: ICLR, 2014.
19	Radford A, Metz L, Chintala S. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks[EB/OL]. (2016-01-07)[2022-07-08]. .
20	Dumoulin Vincent, Visin Francesco. A Guide to Convolution Arithmetic for Deep Learning[EB/OL]. (2018-01-11) [2022-07-22]. .
21	Mirza Mehdi, Osindero S. Conditional Generative Adversarial Nets[EB/OL]. (2014-11-06) [2022-07-15]. .
22	Cover T M, Hart P E. Nearest Neighbor Pattern Classification[J]. IEEE Transactions on Information Theory, 1967, 13(1): 21-27.
23	Goodfellow Ian, Bengio Yoshua, Courville Aaron. Deep Learning[M]. Cambridge: MIT Press, 2016: 106-140.
24	Breiman L. Random Forests[J]. Machine Learning, 2001, 45(1): 5-32.
25	Freund Y, Schapire R E. A Decision-theoretic Generalization of on-line Learning and an Application to Boosting[J]. Journal of Computer and System Sciences, 1997, 55(1): 119-139.
26	Friedman J H. Greedy Function Approximation: A Gradient Boosting Machine[J]. Annals of Statistics, 2001, 29(5): 1189-1232.
27	Chen Tianqi, Guestrin C. XGBoost: A Scalable Tree Boosting System[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York, NY, USA: Association for Computing Machinery, 2016: 785-794.
28	Ke Guolin, Meng Qi, Finley T, et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook, NY, USA: Curran Associates Inc., 2017: 3149-3157.
29	Prokhorenkova Liudmila, Gusev Gleb, Vorobev Aleksandr, et al. CatBoost: Unbiased Boosting with Categorical Features[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. Red Hook, NY, USA: Curran Associates Inc., 2018: 6639-6649.
30	Akiba Takuya, Sano Shotaro, Yanase Toshihiko, et al. Optuna: A Next-generation Hyperparameter Optimization Framework[C]//Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York, NY, USA: Association for Computing Machinery, 2019: 2623-2631.
31	Ozaki Yoshihiko, Tanigaki Yuki, Watanabe Shuhei, et al. Multiobjective Tree-structured Parzen Estimator for Computationally Expensive Optimization Problems[C]//GECCO 2020-Proceedings of the 2020 Genetic and Evolutionary Computation Conference. New York, NY, USA: Association for Computing Machinery, 2020: 533-541.
32	Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-scale Image Recognition[EB/OL]. (2015-04-10) [2022-08-03]. .
33	Gu Jiuxiang, Wang Zhenhua, Kuen J, et al. Recent Advances in Convolutional Neural Networks[J]. Pattern Recognition, 2018, 77: 354-377.
34	Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift[C]//Proceedings of the 32nd International Conference on International Conference on Machine Learning-Volume 37. Cambridge: JMLR, 2015: 448-456.
35	Santurkar S, Tsipras D, Ilyas A, et al. How Does Batch Normalization Help Optimization?[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. Red Hook, NY, USA: Curran Associates Inc., 2018: 2488-2498.
36	Kingma Diederik P, Ba J L. Adam: A Method for Stochastic Optimization[EB/OL]. (2017-01-30) [2022-07-29]. .
37	石洪波, 陈雨文, 陈鑫. SMOTE过采样及其改进算法研究综述[J]. 智能系统学报, 2019, 14(6): 1073-1083.
	Shi Hongbo, Chen Yuwen, Chen Xin. Summary of Research on SMOTE Oversampling and Its Improved Algorithms[J]. CAAI Transactions on Intelligent Systems, 2019, 14(6): 1073-1083.

模型		准确率	召回率	精确率	F1	AUC
LR	训练集	61.5	62.3	32.6	42.8	67.0
LR	测试集	62.1	62.7	33.1	43.3	67.1
DT	训练集	81.5	48.8	62.9	54.9	80.8
DT	测试集	81.5	42.1	65.8	51.4	80.1

算法	原理	优点	缺点
LR	将目标函数值映射到Sigmoid函数，通过极大似然函数进行分类	可解释性好，训练速度快	处理非线性和不平衡样本问题能力弱
KNN	计算样本之间的距离后进行分类	模型结构简单，计算复杂度低^[22]	处理高维特征能力不足
SVM	通过核函数将样本映射到高维空间，采用超平面进行分类	对高维小样本数据学习能力较强	模型精度过分依赖核函数，计算量过大
MLP	在前馈网络结构基础上，应用反向传播算法和激活函数构造的人工神经网络模型	能够处理非线性问题、鲁棒性较强^[23]	受离散值影响较大，可解释性低
DT	根据信息增益方向进行自上而下分类的树结构模型	模型分类规则具有可解释性	处理缺失数据能力弱、容易过拟合
RF	基于Bagging理论和CART树的集成模型^[24]	处理高维数据能力强，对噪声和异常值有较好的容忍性	处理小样本数据和低维数据能力弱、可解释性差
AdaBoost	采用自适应增加预测错误样本权重的方式实现Boosting	模型结构简单，使用灵活，抗过拟合能力强^[25]	易受噪声干扰，训练耗时长
GBDT	采用梯度下降的思想对残差进行优化^[26]	可以灵活地处理多种类型特征	难以并行化处理，计算复杂度高
XGBoost	在GBDT基础上对损失函数进行二阶泰勒展开并添加正则项	支持并行计算，内置交叉验证，模型泛化能力强	调参困难，内存消耗大^[27]
LightGBM	采用直方图、梯度单边采样算法和带深度限制的分支策略^[28]	训练效率高，内存消耗少	对噪声较为敏感
CatBoost	以对称二叉树为基模型，采用防止梯度估计偏差的排序提升方法^[29]	离散型特征，调参成本低，模型精度高，预测快	离散型特征的处理消耗大量的内存和时间

模型		准确率	召回率	精确率	F1	AUC
LR	训练集	0.823	0.500	0.651	0.566	0.859
LR	测试集	0.825	0.503	0.659	0.570	0.856
MLP	训练集	0.839	0.545	0.695	0.611	0.884
MLP	测试集	0.839	0.546	0.694	0.611	0.882
DT	训练集	0.816	0.536	0.686	0.490	0.831
DT	测试集	0.814	0.531	0.612	0.568	0.825
RF	训练集	0.828	0.626	0.629	0.627	0.872
RF	测试集	0.825	0.655	0.613	0.633	0.870
AdaBoost	训练集	0.836	0.537	0.685	0.602	0.880
AdaBoost	测试集	0.835	0.532	0.686	0.599	0.878
GBDT	训练集	0.841	0.528	0.771	0.606	0.877
GBDT	测试集	0.840	0.524	0.708	0.602	0.883
XGBoost	训练集	0.805	0.803	0.553	0.655	0.892
XGBoost	测试集	0.803	0.787	0.552	0.649	0.889
LightGBM	训练集	0.798	0.821	0.541	0.653	0.893
LightGBM	测试集	0.794	0.810	0.537	0.646	0.889
CatBoost	训练集	0.832	0.696	0.623	0.657	0.893
CatBoost	测试集	0.829	0.693	0.616	0.652	0.890

模型		准确率	召回率	精确率	F1	AUC
Voting	训练集	0.824	0.736	0.596	0.658	0.892
Voting	测试集	0.821	0.732	0.591	0.654	0.888
Stacking	训练集	0.843	0.601	0.684	0.640	0.894
Stacking	测试集	0.840	0.595	0.674	0.632	0.891
Blending	训练集	0.845	0.598	0.690	0.641	0.894
Blending	测试集	0.837	0.579	0.672	0.622	0.888
gcForest	训练集	0.836	0.509	0.698	0.588	0.873
gcForest	测试集	0.837	0.515	0.701	0.594	0.874

模型		准确率	召回率	精确率	F1	AUC
Stacking	训练集	0.843	0.603	0.680	0.640	0.894
Stacking	测试集	0.839	0.603	0.674	0.631	0.891
SMOTE-Stacking	训练集	0.891	0.895	0.889	0.892	0.964
SMOTE-Stacking	测试集	0.830	0.666	0.625	0.645	0.886
CGAN- Stacking	训练集	0.877	0.827	0.860	0.843	0.952
CGAN- Stacking	测试集	0.831	0.638	0.651	0.644	0.891
ICGAN- Stacking	训练集	0.872	0.847	0.835	0.841	0.951
ICGAN- Stacking	测试集	0.833	0.644	0.678	0.646	0.909