Parallel Computing for Bayesian Networks - Bayesian Networks in R: With Applications in Systems Biology - page 112

> parallel.multistart = function(net) {
+ tabu(hailfinder, start = net)

+}
> netlist = parLapply(cl, start, parallel.multistart)
> unlist(lapply(netlist, score, data = hailfinder))
[1] -990474.8 -997597.7 -991934 -993547.3

It is important to note that execution time is not reduced by the parallel multistart,
because each of the instances executed by the slave processes takes on average as
much time as the original score-based algorithm.

> s0 = random.graph(names(hailfinder),
+ method = "melancon")
> system.time(tabu(hailfinder, start = s0))
user system elapsed
414.130 0.000 414.137
> system.time(parLapply(cl, netlist,
+ parallel.multistart))

user system elapsed
0.020 0.010 432.221
More advanced approaches and applications are available in literature, each tai-
lored to particular problems and with specific advantages and limitations. For an
extensive coverage of such approaches, we refer the reader to Rauber and Runger
( 2010 ).

5.3.3 Hybrid Structure Learning Algorithms

Applications of parallel computing to hybrid algorithms depend on the exact imple-
mentation of the restrict and maximize phases.
The restrict phase is usually implemented using the first two steps of a constraint-
based algorithm or using another local search algorithm . Some examples of the
latter are proposed in Friedman et al. ( 1999b ) for the Sparse Candidate algorithm.
More complex ones, such as ARACNE ( Margolin et al. , 2006 ), are investigated in
Meloni et al. ( 2009 ). Therefore, all the considerations we made in Sect. 5.3.1 apply.
The maximize phase is usually implemented using a score-based learning algo-
rithm. The computational cost of this phase is reduced by the constraints learned
in the restrict phase, which enforce the sparseness of the network structure. This
in turn guarantees a reasonable performance for most real-world data sets. All the
considerations we made in Sect. 5.3.2 still apply; for example, we can still imple-
ment the multistart model if we take care to select starting networks that satisfy the
constraints.