Accelerating point-based POMDP algorithms via greedy strategies

Abstract

Many planning tasks of autonomous robots can be modeled as partially observable Markov decision process (POMDP) problems. Point-based algorithms are well-known algorithms for solving large-scale POMDP problems. Several leading point-based algorithms eschew some flawed but very useful heuristics to find an ε-optimal policy. This paper aims at exploiting these avoided heuristics by a simple framework. The main idea of this framework is to construct a greedy strategy and combine it with the leading algorithms. We present an implementation to verify the framework's validity. The greedy strategy in this implementation stems from some common ignored heuristics in three leading algorithms, and therefore can be well combined with them. Experimental results show that the combined algorithms are more efficient than the original algorithms. On some benchmark problems, the combined algorithms have achieved about an order of magnitude improvement in runtime. These results provide an empirical evidence for our proposed framework's efficiency. © 2010 Springer-Verlag Berlin Heidelberg.