为了实现逆轨拦截机动目标时达到最大拦截末速的目标，基于序列到序列方法构建目标机动弹道分支预测模型，基于深度Q学习算法和偏置制导律构建深度强化学习智能制导律。对于智能制导律训练过程中引发的稀疏奖励问题，采用预测-校正方法引入末制导的导引比以构建终端奖励，结合物理过程和性能指标构建过程奖励，结合过程奖励与终端奖励以提升训练效果。仿真表明目标机动弹道分支预测模型与弹道外推方法相比，在分方向上的机动弹道平均预测精度至少提高67%；在达到交班性能要求和低过载要求的前提下，智能制导律相比于基线制导律在相对拦截速度上提升67%。

To achieve the maximum interception terminal velocity when intercepting maneuvering targets in the contra-orbit， this paper constructs a target maneuvering ballistic branch prediction model based on the sequence-to-sequence method， and constructs a deep reinforcement learning intelligent guidance law based on the deep Q-learning algorithm and the bias guidance law. To address the sparse reward problem caused by the training process of the smart guidance law， the prediction-correction method is used to introduce the guidance ratio of the terminal guidance to construct the terminal reward， and the process reward is constructed by combining the physical process and performance indexes. The process reward and terminal reward are combined to improve the training effect. Simulation shows that the target maneuvering ballistic branch prediction model improves the average prediction accuracy of the maneuvering ballistic in the sub-direction by at least 67% compared with the ballistic extrapolation method， and the intelligent guidance law improves the relative interception speed by 67% compared with the baseline guidance law on the premise of meeting the requirements of shift performance and low overload.