针对深度Q网络(DQN)算法因过估计导致收敛稳定性差的问题,在传统时序差分(TD)的基础上提出N阶TD误差的概念,设计基于二阶TD误差的双网络DQN算法。构造基于二阶TD误差的值函数更新公式,同时结合DQN算法建立双网络模型,得到两个同构的值...针对深度Q网络(DQN)算法因过估计导致收敛稳定性差的问题,在传统时序差分(TD)的基础上提出N阶TD误差的概念,设计基于二阶TD误差的双网络DQN算法。构造基于二阶TD误差的值函数更新公式,同时结合DQN算法建立双网络模型,得到两个同构的值函数网络分别用于表示先后两轮的值函数,协同更新网络参数,以提高DQN算法中值函数估计的稳定性。基于Open AI Gym平台的实验结果表明,在解决Mountain Car和Cart Pole问题方面,该算法较经典DQN算法具有更好的收敛稳定性。展开更多
离线强化学习通过减小分布偏移实现了习得策略向行为策略的逼近,但离线经验缓存的数据分布往往会直接影响习得策略的质量.通过优化采样模型来改善强化学习智能体的训练效果,提出两种离线优先采样模型:基于时序差分误差的采样模型和基于...离线强化学习通过减小分布偏移实现了习得策略向行为策略的逼近,但离线经验缓存的数据分布往往会直接影响习得策略的质量.通过优化采样模型来改善强化学习智能体的训练效果,提出两种离线优先采样模型:基于时序差分误差的采样模型和基于鞅的采样模型.基于时序差分误差的采样模型可以使智能体更多地学习值估计不准确的经验数据,通过估计更准确的值函数来应对可能出现的分布外状态.基于鞅的采样模型可以使智能体更多地学习对策略优化有利的正样本,减少负样本对值函数迭代的影响.进一步,将所提离线优先采样模型分别与批约束深度Q学习(Batch-constrained deep Q-learning,BCQ)相结合,提出基于时序差分误差的优先BCQ和基于鞅的优先BCQ.D4RL和Torcs数据集上的实验结果表明:所提离线优先采样模型可以有针对性地选择有利于值函数估计或策略优化的经验数据,获得更高的回报.展开更多
Based on error analysis, the influence of error sources on strapdown inertial navigation systems is discussed. And the maximum permissible component tolerances are established. In order to achieve the desired accuracy...Based on error analysis, the influence of error sources on strapdown inertial navigation systems is discussed. And the maximum permissible component tolerances are established. In order to achieve the desired accuracy (defined by circular error probability), the types of appropriate sensors are chosen. The inertial measurement unit (IMU) is composed of those sensors. It is necessary to calibrate the sensors to obtain their error model coefficients of IMU. After calibration tests, the accuracy is calculated by uniform design method and it is proved that the accuracy of IMU is satisfied for the desired goal.展开更多
In this paper,a implicit difference scheme is proposed for solving the equation of one_dimension parabolic type by undetermined paameters.The stability condition is r=αΔt/Δx 2 1/2 and the truncation error is o(...In this paper,a implicit difference scheme is proposed for solving the equation of one_dimension parabolic type by undetermined paameters.The stability condition is r=αΔt/Δx 2 1/2 and the truncation error is o(Δt 4+Δx 4) It can be easily solved by double sweeping method.展开更多
文摘针对深度Q网络(DQN)算法因过估计导致收敛稳定性差的问题,在传统时序差分(TD)的基础上提出N阶TD误差的概念,设计基于二阶TD误差的双网络DQN算法。构造基于二阶TD误差的值函数更新公式,同时结合DQN算法建立双网络模型,得到两个同构的值函数网络分别用于表示先后两轮的值函数,协同更新网络参数,以提高DQN算法中值函数估计的稳定性。基于Open AI Gym平台的实验结果表明,在解决Mountain Car和Cart Pole问题方面,该算法较经典DQN算法具有更好的收敛稳定性。
文摘离线强化学习通过减小分布偏移实现了习得策略向行为策略的逼近,但离线经验缓存的数据分布往往会直接影响习得策略的质量.通过优化采样模型来改善强化学习智能体的训练效果,提出两种离线优先采样模型:基于时序差分误差的采样模型和基于鞅的采样模型.基于时序差分误差的采样模型可以使智能体更多地学习值估计不准确的经验数据,通过估计更准确的值函数来应对可能出现的分布外状态.基于鞅的采样模型可以使智能体更多地学习对策略优化有利的正样本,减少负样本对值函数迭代的影响.进一步,将所提离线优先采样模型分别与批约束深度Q学习(Batch-constrained deep Q-learning,BCQ)相结合,提出基于时序差分误差的优先BCQ和基于鞅的优先BCQ.D4RL和Torcs数据集上的实验结果表明:所提离线优先采样模型可以有针对性地选择有利于值函数估计或策略优化的经验数据,获得更高的回报.
文摘Based on error analysis, the influence of error sources on strapdown inertial navigation systems is discussed. And the maximum permissible component tolerances are established. In order to achieve the desired accuracy (defined by circular error probability), the types of appropriate sensors are chosen. The inertial measurement unit (IMU) is composed of those sensors. It is necessary to calibrate the sensors to obtain their error model coefficients of IMU. After calibration tests, the accuracy is calculated by uniform design method and it is proved that the accuracy of IMU is satisfied for the desired goal.
文摘In this paper,a implicit difference scheme is proposed for solving the equation of one_dimension parabolic type by undetermined paameters.The stability condition is r=αΔt/Δx 2 1/2 and the truncation error is o(Δt 4+Δx 4) It can be easily solved by double sweeping method.