A New Interpolation Scheme in Grid-Based Approximation for POMDPs

Finding a solution to the scalable partially observable Markov decision process (POMDP) has received considerable attention, since the POMDP framework represents an important tool to model a rich variety of real world sequential decision processes. Due to the intractability of computing exact solutions for POMDPs the researchers in the literature are seeking approximation solutions. The approximation by value iterating over a finite set of belief points is know as grid-based method. This paper proposes a special interpolation scheme collecting reachable belief points to form the finite belief set. The approach is based on the observation that reachable beliefs are what we are really interested in other than those arbitrary ones. The experiments on three benchmark problems show that our approach generates very good results.