Generating Exponentially Smaller POMDP Models Using Conditionally Irrelevant Variable Abstraction

摘要

The state of a POMDP can often be factored into a tuple of n state variables. The corresponding flat model, with size exponential in n, may be intractably large. We present a novel method called conditionally irrelevant variable abstraction (CIVA) for losslessly compressing the factored model, which is then expanded into an exponentially smaller flat model in a representation compatible with many existing POMDP solvers. We applied CIVA to previously intractable problems from a robotic exploration domain. We were able to abstract, expand, and approximately solve POMDPs that had up to 10~24 states in the uncompressed flat representation.