Ole J. Mengshoel
Belief networks (BNs) are an essential knowledge representation technique in AI (Pearl 1988). Substantial progress has been made over the last ten years in all areas of BN research. However, at AAAI-96, Horvitz called for more research related to handling time, synchronicity, and streams of events (Selman et el. 1996): "We [...] need to develop better means of synchronizing an agent’s perceptions, inference, and actions with important events in the world." In this abstract, I consider how to achieve efficient and synchronized inference by approximate BN inference and BN approximation. For approximate BN inference I investigate genetic algorithms (GAS). GAS are robust function optimizers that employ stochastic, instance-based (or population-based) search (Goldberg 1989). Since a BN represents a function, it is natural to consider using GAS to search in the space of instantiated BNs, and to combine GA search with existing algorithms for BN inference.