A robust localization algorithm has been proposed for an Unmanned Surface Vehicle (USV) using DDQN-AM (Double Deep Q-Network with Action Memory). Inertial Navigation System (INS) and Global Positioning System (GPS) are used for collecting location information and for controlling the USV with DDQN-AM algorithm. Conventional probability-based filters such as Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) are suitable for estimating behaviors of nonlinear systems and have been widely applied for localizing various kinds of nonlinear vehicles. However, the algorithms diverge or linearization position error increases in localizing USVs that are required to follow a desired trajectory accurately under high wave and wind conditions. Also, computational complexity is very high, which causes the instability of the vehicle system. To address these problems, DDQN with Adam optimization for error back propagation have been adopted in this research. However, if learning is performed using the training data set of the current time (current state, reward, action) in the replay buffer of DDQN, the learning performance (accuracy for states and running time) may deteriorate and learning load may occur. therefore, to improve the learning performance of DDQN, DDQN-AM has been newly proposed by adding AM in this research. To verify the effectiveness and to show the superior learning performance of the proposed DDQN-AM against the conventional approaches, a simulation using Matlab and field experiments with a prototype USV have been performed at the seashore. As a result, the position error of DDQN-AM was improved by 57.6% compared to the existing approach, and the average running time was also shortened by 1 minute, 30 seconds.

已经提出了使用DDQN-AM（带动作记忆的双深度Q网络）的无人水面飞行器（USV）的鲁棒定位算法。惯性导航系统（INS）和全球定位系统（GPS）用于收集位置信息并通过DDQN-AM算法控制USV。诸如扩展卡尔曼滤波器（EKF）和无味卡尔曼滤波器（UKF）之类的基于概率的常规滤波器适用于估计非线性系统的行为，并已广泛应用于各种非线性车辆的定位。但是，在高风浪条件下，为了精确地遵循所需轨迹而定位的USV，算法会发散或线性化位置误差会增加。另外，计算复杂度非常高，这导致车辆系统的不稳定。为了解决这些问题，本研究中采用了带有误差回传的Adam优化的DDQN。但是，如果使用DDQN的重放缓冲区中当前时间（当前状态，奖励，动作）的训练数据集进行学习，则学习性能（状态和运行时间的准确性）可能会降低，并且可能会产生学习负担。因此，为了提高DDQN的学习性能，在本研究中通过添加AM提出了DDQN-AM。为了验证有效性并显示所提出的DDQN-AM相对于传统方法的优越学习性能，已经在海边进行了使用Matlab的仿真和带有原型USV的现场实验。结果，与现有方法相比，DDQN-AM的位置误差提高了57.6％，