Bayesian Network Construction for Contingent Probabilistic Plan Evaluation

Planning with Incomplete Information for Robot Problems

Volume

Planning with Incomplete Information for Robot Problems

Contents

Contingent probabilistic plans (CPP) have emerged to be an important tool for planning in uncertain and incomplete environments. While Bayesian networks have been used as an efficient mechanism for probabilistic plan projection, their use for CPPs evaluation has not been discussed. In this paper, we address not only the procedures to construct compact Bayesian networks for CPPs evaluation but also their correctness, an aspect neglected by most current Bayesian network approaches. We try to accomplish that by providing a language which combines extended logic programs and probabilistic logic programs to represent probabilistic and deterministic actions models with indirect effects and domain constraints. We propose the concept of Bayesian network-trees which are compact tree structures of Bayesian network fragments and use them to evaluate the probability that a CPP achieves a given goal. We discuss Bayesian network abstraction techniques which can be used to speed up the evaluation process. Our Bayesian networktree construction algorithm can be proved to be sound and complete, under certain conditions, in our proposed semantics.