基于注意力-生成对抗网络的任务分析方法研究  

Study on Task Analysis Methods Based on Attention-GAN

作　　者：周琳茹 彭鹏菲[1] ZHOU Linru;PENG Pengfei(School of Electronic Engineering,Naval University of Engineering,Wuhan 430030,China)

机构地区：[1]中国人民解放军海军工程大学电子工程学院,武汉430030

出　　处：《计算机科学》2024年第3期63-71,共9页Computer Science

摘　　要：合理的任务分析可帮助分析者快速、准确地进行任务规划,目前使用案例推理方法进行任务分析存在分析时间长、分析结果准确性较低等问题。针对该问题,提出了基于注意力-生成对抗网络的任务分析方法。以长短时记忆网络(LSTM)为生成器、循环神经网络(RNN)为判别器,针对离散数据细微梯度的更新无法回传的问题,在生成器中使用rollout policy对生成的不完整序列进行推理补充,在判别器中使用蒙特卡罗(MC)进行数据采样得到完整的数据序列动作价值函数,从而指导生成器的参数更新;针对稀疏数据特征不明显、数据重点不明确等问题,在生成对抗网络训练前加入软注意力机制,为不同特征赋予不同权重从而过滤冗余数据,筛选出重要的特征数据。将该方法与未加入注意力机制的生成对抗网络在同一模拟数据集上进行对比实验,结果表明,加入注意力机制后的方法在精确率(P)、召回率(R)、F1值和准确率(Accuracy)4种评价指标上分别提升了0.088,0.092,0.094和0.068,与其他神经网络推荐算法相比,在P,R,F1值和Accuracy上分别提升了0.1~0.3,0.1~0.2,0.1~0.25和0.07~0.17,证明了该方法的有效性。A reasonable task analysis can help decision makers to plan tasks quickly and accurately.The current task analysis method using case-based reasoning has problems such as long analysis time and low accuracy of analysis results.The method uses LSTM as the generator and RNN as the discriminator.For failing to return updates for small gradients of discrete data,the ge-nerator uses rollout policy to complete the incomplete sequence generated and the discriminator uses Monte Carlo(MC)to sample the data to obtain the complete data sequence action value function,thus guiding on updating the parameters of the generator.To address the problems of sparse data with obscure features and unclear data focus,a soft attention mechanism is added to the GAN before training.It assigns different weights to different features to filter out redundant data and select the important features.The proposed method is compared with the GAN without the attention mechanism on the same simulated dataset,and it is demonstrated that the method with the attention mechanism improves 0.088,0.092,0.094 and 0.068 in terms of P,R,F1 value and accuracy respectively.Compared with other neural network recommendation algorithms,P,R,F1 valueand accuracy is improved by 0.1~0.3,0.1~0.2,0.1~0.25 and 0.07~0.17,respectively,which proves the effectiveness of the proposed method.

关 键 词：注意力机制 生成对抗网络 任务分析 循环神经网络 任务推荐 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]