Robust Binary Search provides a binary search implementation which is robust against errors during the search. In other words, if the comparison function sometimes returns an incorrect result, the search in this project will still converge on the correct solution.
This is adapted from the multiplicative weights algorithm in "Noisy binary search and its applications" by Karp and Kleinberg, with adjustments to make it deterministic and then extended to support directed acyclic graphs.
To use the git bisect replacement, install with cargo install robust-git-bisect, and then
~/.cargo/bin/robust-git-bisect $start_commit $end_commit $command_to_test_commit
For library usage, see AutoSearcher for binary search over a linear range and
AutoCompressedDAGSearcher for binary search over a graph.
This code is optimized to minimize the number of tests executed (i.e. number of iterations) and not necessarily the CPU time of the search algorithm itself, so this will be slower than a plain binary search if the test is deterministic.
The linear algorithm (Searcher and AutoSearcher) takes approximately O(log N) time per
iteration. The graph algorithm (CompressedDAGSearcher and AutoCompressedDAGSearcher) takes
approximately O(segments) time per iteration.
robust-git-bisect shows improved performance compared with git bisect (higher accuracy with fewer iterations):
| Method | Iterations | Accuracy |
|---|---|---|
| robust-git-bisect with 0.99 target | 29.6558 | 99.5392% |
| robust-git-bisect with 0.9 target | 26.1828 | 98.8950% |
| git bisect | 16.1907 | 31.7972% |
| git bisect with tests repeated | 35.0465 | 86.6359% |
| git bisect repeated | 72.3674 | 86.1751% |
This test is run over the git git repo from e83c516331 to 54e85e7af1, simulating 9c3592cf3c as the
bad commit, with a test that returns an incorrect result 5% of the time. See benchmark.rs for
details.