摘要
针对现有方法难以预测出符合飞行移动目标不确定行为方式轨迹的问题,提出基于逆强化学习的飞行移动目标轨迹预测方法,通过学习目标行为偏好以及模拟目标行为决策过程的方式预测目标的移动轨迹。首先基于深度神经网络建立目标的行为决策模型与行为偏好模型,然后通过最大熵逆强化学习方法交替地学习模型参数。为了有效地学习目标的不确定行为特征,采用监督学习的方法学习出目标示例轨迹概率分布模型,用于指导目标行为偏好模型的训练以及初始化目标行为决策模型,同时通过对目标行为偏好模型进行预训练的方式提高其训练质量。仿真结果表明,提出的飞行移动目标轨迹预测方法可通过学习到的目标行为决策模型较为准确地模拟目标的行为方式,预测的目标轨迹分布与真实的目标轨迹分布在Kullback-Leibler(KL)散度下的相似度可达0.24。
Aiming at the problem that the existing methods are difficult to predict the trajectory of the flying moving target with uncertain behavior patterns,a trajectory prediction method for flying moving targets based on inverse reinforcement learning is proposed,which can predict the moving trajectory by learning the behavior preference of the target and simulating the decision-making process of the target behavior.Firstly,the behavior decision model and behavior preference model of the target are established based on deep neural networks,and then the model parameters are alternately learned by a maximum entropy inverse reinforcement learning method.In order to effectively learn the uncertain behavior characteristics of the target,the supervised learning method is used to learn the probability distribution model of the target sample trajectories,which are then used to guide the training of the target behavior preference model and initialize the target behavior decision model.Meanwhile,the training quality of the target behavior preference model is improved by pre-training.The simulation results show that the proposed method can accurately simulate the behavior patterns of the target through the learned target behavior decision model,and the similarity between the predicted target trajectory distribution and the real target trajectory distribution under Kullback-Leibler(KL)divergence can reach 0.24.
作者
颜鹏
郭继峰
白成超
YAN Peng;GUO Jifeng;BAI Chengchao(School of Astronautics,Harbin Institute of Technology,Harbin 150001,China)
出处
《宇航学报》
EI
CAS
CSCD
北大核心
2022年第8期1040-1051,共12页
Journal of Astronautics
基金
国家自然科学基金(61973101)。
关键词
轨迹预测
飞行移动目标
不确定行为方式
逆强化学习
深度神经网络
Trajectory prediction
Flying moving targets
Uncertain behavior patterns
Inverse reinforcement learning
Deep neural networks