Prior knowledge for learning networks in non-probabilistic settings

AbstractCurrent learning methods for general causal networks are basically data-driven. Exploration of the search space is made by resorting to some quality measure of prospective solutions. This measure is usually based on statistical assumptions. We discuss the interest of adopting a different point of view closer to machine learning techniques. Our main point is the convenience of using prior knowledge when it is available. We identify several sources of prior knowledge and define their role in the learning process. Their relation to measures of quality used in the learning of possibilistic networks are explained and some preliminary steps for adapting previous algorithms under these new assumptions are presented