基于神经网络的实时滚动追逃博弈导弹制导律

针对导弹实时滚动追逃博弈对抗双方制导律求解问题,设置了若干组对抗双方初始状态,采用分解正交配置法分别离线求解双边开环最优控制,并组成神经网络训练数据集。基于数据集将所有短周期初始和终止时刻对抗双方的状态和控制变量作为输入和输出,采用反向传播(back propagation,BP)算法训练神经网络。然后分别在简单、复杂和不确定环境下,基于滚动时域优化框架使用BP神经网络估计短优化周期内双边开环最优控制,反馈更新对抗双方状态并重复上述过程,进而实时滚动求解导弹追逃博弈双边闭环最优控制。最后将上述方法和直接法得到的优化结果进行比较,捕捉点位置和博弈时间最大误差分别为0.554%和0.097%,两种方法的优化结果吻合较好。同时本文方法计算耗时明显下降,有效提高了导弹滚动追逃博弈制导律求解的实时性。

Real time receding horizon pursuit and evasion games of missile guidance based on neural network

For the problem of real-time receding horizon pursuit and evasion games of missile guidance, this study sets up several groups of initial states of confrontation missiles. Then the decomposing orthogonal collocation method is utilized to solve the bilateral open-loop optimal control of missiles pursuit and evasion games offline, which forms the training data set. The state and control variables of confrontation missiles at the initial and terminal moments are utilized as input and output respectively to train the neural network with the back propagation (BP) algorithm. In the simple, complex and uncertain environment, the BP neural network is utilized to estimate the bilateral open-loop optimal control of missiles in the short optimization cycle based on the receding horizon optimization framework. The states of confrontation missiles are fed back and updated, and the bilateral closed-loop optimal control of missiles pursuit and evasion games is solved in real time. This study compares the optimization results obtained by neural network and direct methods, respectively. The maximum errors of the capture point position and game time are 0.554% and 0.097%, respectively, which show that the optimization results of the two methods agree well. The calculation time of the neural network method is obviously reduced compared with the direct method, which improves the real time solution of receding horizon pursuit and evasion games of missile guidance.