Goal Recognition with Noisy Observations

Goal Recognition with Noisy Observations
复制标题

带有噪声观测的目标识别

DOI:
--
复制
发表时间:
2017
期刊:
--
影响因子:
--
通讯作者:
David E. Smith
David E. Smith
中科院分区:
--
文献类型:
--
作者:
Yolanda E;David E. Smith

文献摘要

被引文献

相似文献

对于涉及代理之间合作的应用程序来说,目标识别是一项重要的技术能力。许多目标识别技术允许观察序列不完整,但很少考虑噪声观察的可能性。在本文中,我们描述了一种基于规划的目标识别方法,该方法可以处理缺失的观测值和观测值中的概率噪声。为此,我们首先使用贝叶斯网络根据观察模型和当前观察序列推断动作概率。然后,我们使用这些信息来估计实现不同可能目标的预期成本。将这些成本与目标的先验成本进行比较可以让我们推断出可能目标的概率分布。
Goal recognition is an important technological capability for applications that involve cooperation between agents. Many goal recognition techniques allow the sequence of observations to be incomplete, but few consider the possibility of noisy observations. In this paper, we describe a planning-based goal recognition approach that deals with both missing observations and probabilistic noise in the observations. To do this, we first use a Bayesian network to infer action probabilities based on the observation model and the current observation sequence. We then use this information to estimate the expected cost of reaching the different possible goals. Comparing these costs to the a priori costs for the goals allows us to infer a probability distribution over the possible goals.