From conventional observation data , it is rarely possible to determine a fully causal Bayesian network. The theoretical point at which we are interested is learning causal Bayesian networks , with or without latent variables. We first focused on the discovery of causal relationships when all variables are known ( ie there are no latent variables ) proposing a learning algorithm using both data from observations and experiments. Logically, we then focused on the same problem when all the variables are not known . We must therefore discover both causal relationships between variables and the presence of latent variables in a Bayesian network structure. To do this, we try to unify two formalisms , semi- Markovian causal models (SMCM) and maximum ancestral graphs (MAG), previously used separately , one for causal inference (SMCM), the other for the discovery of causality (MAG) . We are also interested in the adaptation of causal Bayesian networks for multi -agent systems, and learning these multi-agent causal models (MACM) .
