Abstraction for belief revision: Using a genetic algorithm to compute the most probable explanation

A belief network can create a compelling model of an agent’s uncertain environment. Exact belief network inference, including computing the most probable ex-planation, can be computationally hard. Therefore, it is interesting to perform inference on an approximate belief network rather than on the original belief net-work. This paper focuses on approximation in the form of abstraction. In particular, we show how a genetic al-gorithm can search for the most probable explanation in an abstracted belief network. Belief network approx-imation can be treated as noise from the point of view of a genetic algorithm, and there is therefore a rela-tionship to research on noisy fitness functions used for genetic algorithms