Locally Observable Markov Decision Processes
Locally Observable Markov Decision Processes
复制标题
局部可观察马尔可夫决策过程
DOI：
--
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
G. Konidaris
中科院分区：
文献类型：
--
作者：
Max Merlin;Neev Parikh;Eric Rosen;G. Konidaris
-现实世界中的机器人任务规划在计算上是困难的，部分原因是处理部分可观测性的复杂性。降低规划复杂性的一种方法是假设附加的模型结构，例如混合可观察性、因式分解状态表示或时间扩展的动作。我们引入了一种新的结构化形式，局部Ob可服务的马尔可夫决策过程，它假设部分可观测性源于有限的传感器范围-传感器范围外的对象是不可观测的，但一旦它们在传感器范围内就成为完全可观测的。这种类型的计划求解任务有一个特殊的fic结构：它们必须经过对象从未被观察到完全观察到的位置。我们介绍了一种新的计划器，它通过围绕这些地点构建计划的层次结构来减少计划时间，并将在线计划和fline计划交错。我们给出了一个具有挑战性的领域的初步结果，表明局部性假设使机器人能够在这种类型的不确定性存在的情况下进行有效的规划。
—Real-world robot task planning is computationally intractable in part due to the complexity of dealing with partial observability. One approach to reducing planning complexity is to assume additional model structure such as mixed-observability, factored state representations, or temporally-extended actions. We introduce a novel structured formulation, the Locally Ob-servable Markov Decision Process , which assumes that partial observability stems from limited sensor range—objects outside sensor range are unobserved, but become fully observed once they are within sensor range. Plans solving tasks of this type have a specific structure: they must necessarily go through localities where objects transition from unobserved to fully observed. We introduce a novel planner that reduces planning time via a hierarchy that structures the plan around these localities, and interleaves online and offline planning. We present preliminary results in a challenging domain that shows that the locality assumption enables robots to plan effectively in the presence of this type of uncertainty.