Incremental Learning and Memory Consolidation of Whole Body Motion Patterns
Incremental Learning and Memory Consolidation of Whole Body Motion Patterns
复制标题
DOI：
--
复制
发表时间：
2008
影响因子：
4.7
通讯作者：
Yoshihiko Nakamura
中科院分区：
文献类型：
--
作者：
Yoshihiko Nakamura
在连续和在线观察中学习的能力对人形机器人来说是有利的，因为它将使它们能够在人类环境中的共同定位和互动中学习。然而，当动作被学习和在线聚集时，有一个交易！在分类精度和训练样本数量之间，导致在运动和层次结构形成层面上潜在的错误分类。本文提出了一种快速在线增量学习的方法，结合增量记忆巩固过程纠正最初的错误分类和组织错误，以提高学习动作的稳定性和准确性，类似于在人类运动学习后观察到的记忆巩固过程。在初始组织之后，随机选择运动进行重新分类，在层次结构的低级别和高级别。如果找到更好的重分类，则对知识结构进行重组以符合要求。在包含各种全身运动的运动数据库的增量获取过程中验证了该方法。
The ability to learn during continuous and on-line observation would be advantageous for humanoid robots, as it would enable them to learn during co-location and interaction in the human environment. However, when motions are being learned and clustered on-line, there is a tradeo! between classification accuracy and the number of training examples, resulting in potential misclassifications both at the motion and hierarchy formation level. This paper presents an approach enabling fast on-line incremental learning, combined with an incremental memory consolidation process correcting initial misclassifications and errors in organization, to improve the stability and accuracy of the learned motions, analogous to the memory consolidation process following motor learning observed in humans. Following initial organization, motions are randomly selected for re-classification, at both low and high levels of the hierarchy. If a better reclassification is found, the knowledge structure is re-organized to comply. The approach is validated during incremental acquisition of a motion database containing a variety of full body motions.