Planning for deterministic and probabilistic domains differ significantly in representations they require, the algorithms that solve them and the way in which results are represented. Hierarchical Task Networks (HTN) and Markov Decision Process (MDPs) are representative formalisms of, respectively, deterministic and stochastic planning. Stochastic domain specifications can easily become opaque to a human designer, especially as the domain size increases. Our research aims to develop algorithms for lossless and automatic mapping of HTN models that are easily intelligible to humans into MDPs. In this paper we develop algorithms to convert deterministic planning domains with HTN domain knowledge and an action error model into MDPs that can then be solved, while maintaining a bound on the number of MDP states.
|Title of host publication||Decision Making in Partially Observable, Uncertain Worlds: Exploring Insights from Multiple Communities|
|Number of pages||6|
|Publication status||Published - 2011|