摘要
对战场目标战术意图的快速、准确和自动识别,是智能决策的前提和基础。目标战术意图通常由多个战术动作组合完成,因而目标状态呈现动态、时序变化特征。本文针对目标意图识别问题的特点,提出一种基于栈式自编码器(SAE)的智能识别模型,设计智能识别模型的基本框架,提出一种基于时序特征的输入信号编码方法及相应的模式解析机制,通过将目标状态在多个时刻的时序特征和战场环境、目标属性等信息统一编码为输入信号,将军事专家的知识经验封装为模式标签,模拟人的推理模式与认知经验,实现对目标战术意图的智能识别。最后通过实验,分析预训练过程和网络深度对算法性能的影响,并通过与多层感知机(MLP)和逻辑回归分类器(LRC)识别准确率的比较,验证所提SAE算法的有效性。
Automatic and fast intention recognition is the premise and bedrock of intelligent decision-making, and it refers to the process of deducing the intention of a target from a set of observed actions with dynamic and temporal characteristics.Here, an automatic tactical intention recognition model based on deep learning methods of stacked auto-encoder (SAE) is proposed. The temporal features and attributes of corresponding target and battlefield environment are encoded as the input signal, and then the recognition experience of domain expert are encapsulated and labeled as know-ledge to train the intelligent model so as to simulate the deducing and cognition mode of human. Finally, the influence of pre-training and the depth of the SAE network on the perfor^nance are analyzed, and the validation of SAE model was illustrated by comparison of recognition accuracy ratios with that obtained by the models based on multi-layer perceptron ( MLP) and logistic regression classifier (LRC).
作者
欧微
柳少军
贺筱媛
郭圣明
OU Wei;LIU Shao-jun;HE Xiao-yuan;GUO Sheng-ming(Department of Information Operation & Command Training,NDU, Beijing 100091;Urumqi Border Cadre Training Unit, Urumqi 830002, China)
出处
《指挥控制与仿真》
2016年第6期36-41,共6页
Command Control & Simulation
基金
国家自然科学基金(60403401
61374179
61273189
61174156
61174035)
全军军事学研究生课题(2015JY035)
关键词
意图识别
时序特征
栈式自编码器
深度学习
tactical intention recognition
temporal features
stacked auto-encoder
deep learning
