摘要
针对基于深度学习的作战效能评估领域存在需要人工调参的问题,提出使用Adam优化器和粒子群算法(particle swarm optimization,PSO)进行自动参数寻优的方法。首先,以基于深度置信网络(deep belief network,DBN)的非协作式敌我识别(non-cooperative identification of friend or foe,NC-IFF)系统作战效能评估模型为基础,构建了PSO-DBN、Adam-DBN、Adam-PSO-DBN 3个模型;然后,分别设置损失值达到0.01、训练轮数达到100、程序运行5 min 3种情况,对DBN、PSO-DBN、Adam-DBN、Adam-PSO-DBN进行对比实验;最后,在各模型都完成充分训练的基础上进行实例验证。结果表明:DBN、PSO-DBN、Adam-DBN、Adam-PSO-DBN 4个模型评估结果的均方根误差分别为0.1842、0.1772、0.1829、0.1726;通过引入PSO和Adam优化器可以提升计算效率,优化DBN的训练效果,其中,Adam在短时间内提升效果明显,而PSO会在寻优过程中花费较多时间,但使用寻优所得参数后可以完成反超;在长时间持续训练后,新一轮训练带来的提升不再明显,PSO的优势会减弱,Adam和PSO共同作用时,优化效果相对最好。
To solve the problem of manual parameter tuning in the field of combat effective-ness evaluation based on deep learning,an automatic parameter optimization method using Adam optimizer and particle swarm optimization(PSO)was proposed.First,based on the combat ef-fectiveness evaluation model of the non-cooperative identification of friend or foe(NC-IFF)sys-tem using a deep belief network(DBN),three models were constructed,namely PSO-DBN,Adam-DBN and Adam-PSO-DBN.Comparative experiments were conducted among DBN,PSO-DBN,Adam-DBN,and Adam-PSO-DBN,with the loss value set to 0.01,the number of train-ing rounds set to 100,and the running time of program for 5 min.Finally,an example verifica-tion was carried out on the fact that each model had completed sufficient training.The experi-mental results show that the root mean square errors of the evaluation results for the four mod-els,DBN,PSO-DBN,Adam-DBN and Adam-PSO-DBN are 0.1842,0.1772,0.1829 and 0.1726,respectively.The computing efficiency and training effect of DBN can be improved by in-troducing the PSO and Adam optimizers.Adam can achieve significant improvements in a short time,whereas PSO requires more time for optimization.However,by using the optimized param-eters,PSO can surpass other methods in performance.After a long period of continuous training,the improvement made by the new round of training is no longer evident,and the advantage of PSO is weakened.The best optimization effect is achieved when Adam and PSO are used in combination.
作者
刘赟
张琪
杨立浩
LIU Yun;ZHANG Qi;YANG Lihao(Rocket Force University of Engineering,Xi’an 710025,China)
出处
《火箭军工程大学学报》
2025年第4期124-136,共13页
Journal of Rocket Force University of Engineering
