在面对复杂未知的环境时,智能体能进行大规模探索一直是深度强化学习的研究热点之一,但是传统的深度Q网络采用ε-greedy局部扰动策略来进行探索,这种策略方法不能保证一定学习到有效合理的ε,以达到探索的最优,其次仅根据状态值函数选...在面对复杂未知的环境时,智能体能进行大规模探索一直是深度强化学习的研究热点之一,但是传统的深度Q网络采用ε-greedy局部扰动策略来进行探索,这种策略方法不能保证一定学习到有效合理的ε,以达到探索的最优,其次仅根据状态值函数选择动作并不会引起策略的改变,不能达到深度探索的目的。为了解决该问题,在深度Q网络的全连接层中注入噪声参数,利用带探索性的噪声进行深度探索以弥补传统策略探索的低效性。噪声来自高斯噪声分布,通过方差驱动探索,使得智能体可以发现大量新状态,提供更加丰富的样本,为决策提供有效信息。最终提出一种基于动作空间噪声的深度Q网络模型(Deep Q-Network Based on Action Space Noise)。实验仿真结果表明,和传统的深度Q网络比较,该网络模型在Open AI Gym平台上的部分战略性游戏取得更好的奖赏值。展开更多
Aming at water conservancy project visualization, a hidden-removal method of dam perspective drawings is realized by building a hidden-removal mathematical model for overlapping points location to set up the hidden re...Aming at water conservancy project visualization, a hidden-removal method of dam perspective drawings is realized by building a hidden-removal mathematical model for overlapping points location to set up the hidden relationship among point and plane, plane and plane in space. On this basis, as an example of panel rockfill dam, a dam hidden-removal perspective drawing is generated in different directions and different visual angles through adapting VC++ and OpenGL visualizing technology. The results show that the data construction of the model is simple which can overcome the disadvantages of considerable and complicated calculation. This method also provides the new means to draw hidden-removal perspective drawings for those landforms and ground objects.展开更多
文摘在面对复杂未知的环境时,智能体能进行大规模探索一直是深度强化学习的研究热点之一,但是传统的深度Q网络采用ε-greedy局部扰动策略来进行探索,这种策略方法不能保证一定学习到有效合理的ε,以达到探索的最优,其次仅根据状态值函数选择动作并不会引起策略的改变,不能达到深度探索的目的。为了解决该问题,在深度Q网络的全连接层中注入噪声参数,利用带探索性的噪声进行深度探索以弥补传统策略探索的低效性。噪声来自高斯噪声分布,通过方差驱动探索,使得智能体可以发现大量新状态,提供更加丰富的样本,为决策提供有效信息。最终提出一种基于动作空间噪声的深度Q网络模型(Deep Q-Network Based on Action Space Noise)。实验仿真结果表明,和传统的深度Q网络比较,该网络模型在Open AI Gym平台上的部分战略性游戏取得更好的奖赏值。
基金Supported by National Natural Science Foundation of China (50779005)
文摘Aming at water conservancy project visualization, a hidden-removal method of dam perspective drawings is realized by building a hidden-removal mathematical model for overlapping points location to set up the hidden relationship among point and plane, plane and plane in space. On this basis, as an example of panel rockfill dam, a dam hidden-removal perspective drawing is generated in different directions and different visual angles through adapting VC++ and OpenGL visualizing technology. The results show that the data construction of the model is simple which can overcome the disadvantages of considerable and complicated calculation. This method also provides the new means to draw hidden-removal perspective drawings for those landforms and ground objects.