Learning Probabilistic Action Models from Interpretation Transitions
Learning Probabilistic Action Models from Interpretation Transitions
复制标题
DOI：
--
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
D. Martínez;Tony Ribeiro;Katsumi Inoue;G. Alenyà;C. Torras
中科院分区：
文献类型：
--
作者：
D. Martínez;Tony Ribeiro;Katsumi Inoue;G. Alenyà;C. Torras
近年来，概率规划社区有了很大的进步，规划者现在可以为非常复杂的概率任务提供解决方案。然而，规划者需要有一个代表系统动态的模型，通常这些模型都是手工构建的。在这篇文章中，我们提出了一个从动态系统的状态转移观测中自动推断概率模型的框架。我们提出了对以前从口译转换中学习的工作的扩展。这些工作将一组状态转移作为输入，并构建了实现给定转移关系的逻辑程序。在这里，我们将该方法扩展到学习一组紧凑的概率规划算子，这些算子捕捉到了概率动态。最后，对所学习模型的质量进行了实验验证。
There have been great advances in the probabilistic planning community during recent years, and planners can now provide solutions for very complex probabilistic tasks. However, planners require to have a model that represents the dynamics of the system, and in general these models are built by hand. In this paper, we present a framework to automatically infer probabilistic models from observations of the state transitions of a dynamic system. We propose an extension of previous works that perform learning from interpretation transitions. These works consider as input a set of state transitions and build a logic program that realizes the given transition relations. Here we extend this method to learn a compact set of probabilistic planning operators that capture probabilistic dynamics. Finally, we provide experimental validation of the quality of the learned models.