Avraham Adler — written Mar 14, 2014 — source

There is a lot of literature and debate on how to rank candidates under preferential voting systems. Two of the methods used to determine winners are those based on some form of Borda count and those based on some form of Condorcet method. Many students of politics and voting systems prefer Condorcet methods to Borda ones for its stronger theoretical qualities. However, while Borda counts, especially in their most recognizable form—average rank, is easy to calculate, Condorecet winners are more difficult, as it requires pairwise comparisons between all candidates. Moreover, the possibility of having a Condorcet cycle grows as the number of candidates grows.

One of the more common methods of solving a Condorcet paradox is the Schulze method. This not only has some strong theoretical qualities, but it also has a relatively simple implementation. However, it is slow. Pairwise ranking, in its straightforward form, is of O(n²) and the Schulze method is O(n³) where n is the number of candidates.

In this post, Rcpp is used to significantly speed up the vote ranking process.

Here is a sample ballot with eight voters ranking five candiates:

Candidates Vote A Vote B Vote C Vote D Vote E Vote F Vote G Vote H 1 Albert 1 2 1 3 2 1 3 4 2 Bruce 2 4 5 4 5 4 1 2 3 Charles 3 1 3 1 1 2 4 5 4 David 4 5 2 5 4 5 2 1 5 Edward 5 3 4 2 3 3 5 3

Here is some simple code to calculate the average rank of the candidates:

The above ballot would result in the following Borda-based ranking:

Names Average Rank Position 1 Albert 2.125 1 3 Charles 2.500 2 2 Bruce 3.375 3 4 David 3.500 4 5 Edward 3.500 5

Here is some simplified code to calculate Condorcet and Schulze winners. The ballots have been created to ensure that there is always a unique Schulze winner. In reality, there often is not, and some further form of tiebreaking routine will be necessary:

Using the sample ballot, the pairwise matrix is:

[,1] [,2] [,3] [,4] [,5] [1,] 0 6 5 6 6 [2,] 2 0 3 5 3 [3,] 3 5 0 5 7 [4,] 2 3 3 0 4 [5,] 2 5 1 4 0

the beatpath matrix for the top ranked candidate (using all ballots) is:

[,1] [,2] [,3] [,4] [,5] [1,] 0 6 5 6 6 [2,] 0 0 0 5 0 [3,] 0 5 0 5 7 [4,] 0 0 0 0 0 [5,] 0 5 0 5 0

and the full Condorcet ranking is:

Name Rank Method 1 Albert 1 Condorcet 2 Charles 2 Condorcet 3 Edward 3 Schulze 4 Bruce 4 Condorcet 5 David 5 Condorcet

When profiling this code using the actual ballot of 30+ people with multiple Condorcet paradoxes, 81% of the time was spent in the Schulze algorithm, another 12% was spent in the PairCount algorithm, and the remaining 7% was spent on everything else (the actual ranking had multiple steps to handle cases when there was no Schulze winner). To speed up the procedure, I ported the Schulze and PairCount functions to C++:

It is also interesting to compare these results with those obtained from byte-compiling the functions:

First, we need to check that the functions return the same values:

[1] TRUE

[1] TRUE

Now let’s compare the speed:

Unit: microseconds expr min lq median uq max neval PairCount(V) 375.617 384.634 389.130 400.919 1280.33 100 PairCount_cmp(V) 255.012 263.320 267.079 272.044 361.06 100 PairCount_cmp3(V) 247.473 254.373 259.001 265.401 1120.02 100 PairCount_C(V) 7.334 11.793 12.558 13.214 32.37 100 Schulze(P) 786.678 805.497 812.786 842.798 1677.24 100 Schulze_cmp(P) 230.732 237.673 241.639 251.459 1062.25 100 Schulze_cmp3(P) 195.612 200.953 205.156 214.832 1037.79 100 Schulze_C(P) 3.496 5.712 7.485 8.086 14.26 100

While byte-compiling the PairCount function gives an impressive speedup of
ariund 40%, porting it to C++ makes it over 25 times faster (an over 2400%
speedup). Results with the Schulze algorithm is even more
striking. Byte-compilation gives an increase in speed of between 3 to 3.5
times without any change to the R code, but porting it to C++ is **around 100
times** as fast (and was over 120 times as fast on a different machine)!
Moreover, the PairCount algorithm reads more logically in C++, as the way R
handles vectors and matrices, when subtracting the current rank, the
resulting matrix ended up transposed with the candidates across the columns.

So with easy-to-read code that results in speed gains of multiple orders of magnitude, what’s not to like?!

**tags:**
basics
modeling
benchmark

- Working with Rcpp::StringVector — Brian J. Knaus
- Optimizing Code vs Recognizing Patterns with 3D Arrays — James Joseph Balamuta
- Hierarchical Risk Parity Implementation in Rcpp and OpenMP — Rafael Nicolas Fermin Cota, Yi King, and Chris Chung