逻辑马尔可夫决策编程和关系马尔可夫决策编程若干最新进展

逻辑马尔可夫决策过程和关系马尔可夫决策过程的引入，使得人们可能简洁地、陈述地表达复杂的马尔可夫决策过程。本文首先介绍有关逻辑马尔可夫决策过程和关系马尔可夫决策过程的概念，然后重点介绍它们与普通的马尔可夫决策过程根本不同的一些算法：①依赖于基本状态空间RL的转换法；②把Bellman方程推广到抽象状态空间的方法；③利用策略偏置空间寻求近似最优策略方法。最后对它们的研究现状进行总结及其对它们发展的一些展望。

Several New Advances of Logical Markov Decision Processes and Relational Markov Decision Processes

Using logical Markov decision processes (LDMDPs) and relational Markov decision processes (RMDPs) one can compactly and declaratively represent complex Markov decision processes.This paper firstly introduces central con- cepts of LOMDPs and RMDPs.Then several algorithms that are different from regular Markov decision processes are reviewed:1.The transition method relying on ground state space.2.A relational upgrade of the Bellman update oper- ation.3.Approximate policy iteration using policy bias space.Finally,the paper gives conclusions of recent work and suggests future work.In this way,people are able to get an intensive,comprehensive and in-depth understanding of LOMDPs and RMDPs.