deal (version 1.2-8)

learn: Estimation of parameters in the local probability distributions

Description

Updates the distributions of the parameters in the network, based on a prior network and data. Also, the network score is calculated.

Usage

learn (nw, df, prior=jointprior(nw),
               nodelist=1:size(nw),
               trylist=vector("list",size(nw)),
               timetrace=FALSE)

Arguments

nw

an object of class network.

df

a data frame used for learning the network, see network.

prior

a list containing parameter priors, generated by jointprior.

nodelist

a numeric vector of indices of nodes to be learned.

trylist

a list used internally for reusing learning of nodes, see maketrylist.

timetrace

a logical. If TRUE, prints some timing information on the screen.

Value

• A list with two elements that may be accessed using getnetwork and gettrylist. The elements are
• nw
  an object of class network, with the condposterior properties updated for the nodes. Also, the property score is updated and contains the network score. The contribution to the network score for each node is contained in the property loglik for each node.
• trylist
  an updated list used internally for reusing learning of nodes, see maketrylist.

Details

The procedure learn determines the master prior, local parameter priors and local parameter posteriors, see B�ttcher (2001). It may be called on all nodes (default) or just a single node. From the joint prior distribution, the marginal distribution of all parameters in the family consisting of the node and its parents can be determined. This is the master prior, see localmaster. The local parameter priors are now determined by conditioning in the master prior distribution, see conditional. The hyperparameters associated with the local parameter prior distribution is attached to each node in the property condprior. Finally, the local parameter posterior distributions are calculated (see post) and attached to each node in the property condposterior. A so-called trylist is maintained to speedup the learning process. The trylist consists of a list of matrices for each node. The matrix for a given node holds previously evaluated parent configurations and the corresponding log-likelihood contribution. If a node with a certain parent configuration needs to be learned, it is checked, whether the node has already been learned. The previously learned nodes are given as input in the trylist parameter and is updated in the learning procedure. When one or more nodes in a network have been learned, the network score is updated and attached to the network in the property score. The learning procedure is called from various functions using the principle, that networks should always be updated with their score. Thus, e.g. drawnetwork keeps the network updated when the graph is altered.

References

B�ttcher, S.G. (2001). Learning Bayesian Networks with Mixed Variables, {\emArtificial Intelligence and Statistics 2001}, Morgan Kaufmann, San Francisco, CA, USA, 149-156. Further information about deal can be found at: http://www.math.auc.dk/novo/deal.

See Also

networkfamily, jointprior, maketrylist, network

Examples

data(rats)
fit       <- network(rats)
fit.prior <- jointprior(fit,12)
fit.learn <- learn(fit,rats,fit.prior,timetrace=TRUE)
fit.nw    <- getnetwork(fit.learn)
fit.learn2<- learn(fit,rats,fit.prior,trylist=gettrylist(fit.learn),timetrace=TRUE)