Apprentissage des reseaux Bayesiens causaux a partir de donnees d'observation et d'experimentation
In this paper we discuss the problem of learning Causal Bayesian Networks from a mixture of data and experiments. Any structure learning algorithm, scoring-based or constraint-based, can only learn a Bayesian Network up to Markov equivalence. To find the correct complete structure experiments are needed. We start with a discussion on our notion of what it is to perform experiments or take actions. Next, we study some issues of this experiment phase such as, which experiments to perform and in what order. We allow the possibility to assign a cost and an importance value to each action making some experiments preferred to others. In the case that there are some impossible actions we want to see what the maximal structure is we can retrieve.