Fig. 2

Advising accuracy within benchmark bins. This bar chart shows results within each bin of benchmarks, where bins group benchmarks by difficulty, for parameter advising with greedy advisor sets of cardinality 7. For each of the ten bins listed along the horizontal axis, the vertical axis gives advising accuracy, averaged over the benchmarks in the bin; on the far right is an average of all ten bin-averages. Bins are labeled on the horizontal axis by the upper limit of their difficulty range, where the difficulty of a benchmark is the true accuracy of its alignment computed by the Opal aligner under its default parameter setting. The colored bars in each bin show average advising accuracy for Opal using: its optimal default parameter setting, in black; advising with the original unaugmented Facet estimator, in green; and advising with Facet augmented by predicted coreness, in red. The dashed line shows average advising accuracy for each bin if Facet were augmented by true coreness: the limit achieved by a perfect coreness predictor. In parentheses above the bars is the number of benchmarks in each bin (while on the far right is their total number)