
Software for Bayesian Networks 135
burn-in and start c ollecting MCMC samples. Not knowing the shape of the
posterior distribution, we are not able to assess whether convergence has
been attained or not. Only rough diagnostics are available. In some cases,
several weeks of computations are needed to obtain similar results in JAGS
and OpenBUGS!
• MCMC algorithms are iterative, so they need some initial values from which
to start; this is called the initialisation of the algorithm. BUGS can generate
them automatically; but sometimes this process fails with an error message
like the following.
Erreur dans jags.model(file = "pest.jam", data = dat1) :
Error ...