We implemented and studied a variety of Bayes'-network search
algorithms and scoring metrics. Our experiments allowed the comparison
of their relative performance, and assessment of the relative quality
of the results achieved by the search algorithm variants. We observed
that first-ascent structure search provides approximately the same
result quality as greedy search with considerably faster runtime, but
has increased variance of result quality due to the nondeterminism of
the algorithm. We found that the BIC scoring metric was preferable to
the AIC metric for structure search, since its larger structural
penalty led to less complex network structures. Additionally, analysis
of our experimental results led to a deeper appreciation for the
issues involved in structure search, and the tradeoffs which must be
made when choosing a scoring metric and a search algorithm.