A Belief Network (BN) is a graphical representation of a joint probability distribution over a set of domain variables. Large BNs which model real-time processes are hard to evalaute because of the computational expense. Anytime incremental evaluation algorithms are suitable in such cases. We present a method for anytime evaluation of a BN. Evaluation is initialy performed on a restricted number of nodes in the immediate vicinity of the query nodes. The BN is then travered radially out from each query node and estimates for the blief of the latter are computed interatively. We use a best-first graph traversal strategy to visit in priority the most important nodes while making a trade-off with computation cost. We use arc weights in a BN to determine the effect of a node on the query node, and we also consider the computation cost of visiting a node.
展开▼
