Our experiments demonstrate the applicability of the method. For moderate-size networks the
running times were found to be feasible. The results also highlight the crucial role of the maximum
in-degree and the size of the data (i.e., the polynomial contribution) in the overall complexity.
We also showed that structure discovery on large networks is practical if one can judge a suitable
layering constraint on the space of network structures. Another main conclusion is that the actual
bottleneck in the current implementation is the space requirement, not the speed. For example,
a simple extrapolation (not shown) based on the speed measurements (shown in Table 1) reveals
that a posterior maximizing structure for the Alarm data (37 variables, 2500 examples, maximum
in-degree 4) could be found in about three days, if enough memory is present.