基于环境状态分布优化的POMDP值迭代求解算法

基于点的值迭代算法是一类解决POMDP问题的有效算法，PBVI是基于点集的经典算法，但是其算法效率较为低下。FSVI使用内在的MDP最优策略来降低算法复杂度，但求解大规模问题的效果较差。为解决上述问题，提出了基于环境状态分布优化的前向搜索值迭代算法（PBVI-OSD）,通过基于权重值的Q^MDP选出最佳的动作，基于信念状态和转换函数选取最大可能的状态，基于动作和状态从观察中随机选取一个观察概率大于阈值的观察，由此获得更具探索价值的后继信念点集，提升值迭代收敛的质量。在四个基准问题上的实验表明，相比于FSVI和PBVI,PBVI-OSD能保证收敛效率，特别是在大规模问题上能收敛到更好的全局最优解。

Probability-based value iteration on optimal state distribution algorithm for POMDP

Point-based value iteration methods are a class of practical algorithms for solving the POMDP model. PBVI is a classical point-based value iteration method with low efficiency. FSVI can reduce the complexity and improve efficiency significantly by using the optimal strategy of the underlying MDP. However, it is not efficient in the large-scale POMDP problems. The paper proposed a probability-based value iteration on optimal state distribution algorithm for POMDP(PBVI-OSD). During the exploration, PBVI-OSD used the alias method to sample the action a* based on weighted QMDP function and sample the state based on b and the transition function. Then PBVI-OSD selected the observation z whose probability was greater than the threshold, and got the successor point b with great value from a* and z, which ensured the effect of value iteration. Experiment results of four benchmarks show that PBVI-OSD can achieve better global optimal solutions than FSVI and PBVI, especially in large-scale problems.