Online Resolution Techniques
Online Resolution Techniques
复制标题
在线解决技术
DOI：
10.1002/9781118557426.ch6
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
Frédérick Garçia
中科院分区：
文献类型：
--
作者：
L. Péret;Frédérick Garçia
在马尔可夫决策过程（MDP）的框架中，用于在线确定当前动作的算法通常非常简单。Kearns、Mansour和Ng提出了一种算法，为联合收割机树搜索和在线模拟相结合奠定了理论基础。本章形式化的问题的在线搜索MDP：这个问题可以想象为本地解决方案的MDP在一个给定的推理地平线。它列出了一些解决这个问题的在线方法。与离线算法相比，其主要思想是围绕当前状态进行计算。在线阶段的目的是为每个遇到的状态执行非平凡计算。本章讨论的算法侧重于计算给定当前状态下的最佳操作。它还形式化了在给定的推理范围内确定当前状态的最佳动作的局部问题。
In the framework of Markov decision process (MDP), the algorithm used to determine the current action online is generally very simple. Kearns, Mansour and Ng have proposed an algorithm laying the theoretical foundations to combine tree search and simulation online. This chapter formalizes the problem of the online search for MDP: this problem can be envisioned as the local resolution of an MDP over a given reasoning horizon. It lists some online approaches that solve this problem. Compared to offline algorithms, the main idea is to focus computations around the current state. The purpose of an online phase is to perform a non‐trivial computation for each encountered state. The chapter discusses algorithms that focus on the computation of an optimal action for a given current state. It also formalizes the local problem of determining the best action for a current state over a given reasoning horizon H.Controlled Vocabulary Termstree searching