随机折扣部分可观察马尔可夫决策规划

PARTIALLY OBSERVABLE MARKOV DECISION PROGRAMMING WITH STOCHASTIC DISCOUNT FACTORS

一、引言考虑一个部分可观察马尔可夫决策规划模型(简记为 POMDP),并引文[1]中所用的定义、记号(只将δ_N 换成∏,(?)换成 A)和有关结论.于是信息向量 π(t)与 π(t+1)

The paper discusses an optimal problem for partially observable Markov decision pro-gramming with stochastic variable discount factor over a fintie hirizon.It is shown that theoptimal return function is piecewise-linear and convex and,furthermore,is total-convex.Itis also shown that there exist optimal decision functions that are piecewise-constant.Basedon these results,the “one pass”[1] can be used that only need a few revision.The algorithmis simple and efficient.