Moises Goldszmidt and Adnan Darwiche
This paper describes part of Rockwell’s contribution to Phase I of ARPA’s Planning Initiative (ARPI). Specifically, we describe the representation language and a set of algorithms that constitute the core of a Plan Simulation and Analysis tool (PSA). The main objective of the PSA is to provide capabilities for the testing and evaluation of sequences of actions in domains characterized by unavoidable uncertainties and difficult trade-offs between resources and objectives. The representation language, called action networks, is a semantically well founded framework for reasoning about actions mid change under uncertainty based on probabilistic Bayesian nctworks. Action networks add primitives to Bayesian networks to represent canonical models of time-dependencies, and controllable variables to represent agents manipulations of the domain. In addition, action networks allow different methods for quantifying the uncertainty in causal relationships, which go beyond traditional probabilistic quantification. Inferences are performed via a set of algorithms for approximate computation of belief update. These algorithms allow the user to trade-off computational time for accuracy of the answer, and can be also applied in an anytime fashion.