The Effects of Sampling Kurtosis on Anytime Algorithm Performance

Agent-system and agent-simulator networks encompass complex interactions that are often difficult to model. We use an On-Line Maintenance Agent test bed for analyzing the behavior of Bayesian anytime inference algorithms for influence diagrams. This test bed is used to analyze the performance of a backward simulation algorithm used by the agent for decision making. We show that the kurtosis of sampling acts as a focusing mechanism which may be useful in designing better anytime algorithms.