Embedded bayesian networks: Anyspace, anytime probabilistic inference

An important aspect of probabilistic inference in embedded real-time systems is flexibility to handle changes and limi-tations in space and time resources. We present algorithms for probabilistic inference that focus on simultaneous adapta-tion with respect to these resources. We discuss techniques to reduce memory consumption in Bayesian network inference, and then develop adaptive conditioning, an anyspace anytime algorithm that decomposes networks and applies various al-gorithms at once to guarantee a level of performance. We briefly describe adaptive variable elimination, an anyspace algorithm derived from variable elimination. We present tests and applications with personal digital assistants and industrial controllers