Solving a matching problem with OR-Tools CP-SAT

Thursday March 16, 2023

Optimizing matching for preference satisfaction often focuses on stability, which prioritizes individual interests. When maximizing resource utilization is more important, Constraint Programming (CP) can be used to find good matches. Google's OR-Tools package includes the CP-SAT solver, which is one way to implement this.

For example, we may be assigning a pool of interns to specific roles. Managers for each role have ranked some of the candidates, and a candidate can only be matched to a role for which they have been ranked. Roles often have one spot to be filled, but some have more. Candidates have been ranked for multiple roles, but can only be assigned one.

This problem is similar to the medical residency matching problem, which famously applies the Nobel-prize-winning Gale–Shapley algorithm. One difference is that the present example only has preferences from one side. Sometimes random preferences are created in order to still allow the application of algorithms like Gale–Shapley, but this will not be necessary here because there is a bigger issue.

The salient difference between what Gale–Shapley does and the goal of the example here is that we really want to fill every available role, even if it means the match isn't perfectly stable. Gale–Shapley will create a match that leaves some roles unfilled if this is necessary to ensure that no trade would better satisfy the preferences of the individuals involved. In general, Gale–Shapley will leave some roles unfilled.

To begin a solution with the OR-Tools (OR for Operations Research) package, we start by converting the rankings to scores, where 1 is the best and 0 is the worst: rankings.csv (There are different ways of doing this, like percentiles; here we assume we have scores.)

Role 23,Person 11,0.5238095238095238
Role 23,Person 109,0.7619047619047619
Role 23,Person 111,0.6190476190476191

Each role has some maximum number of spots to fill: spots.csv

Role 23,4
Role 31,1
Role 8,1

We can start reading in data using Python.

import csv

spots = {role: int(n) for role, n in csv.reader(open('spots.csv'))}
all_rankings = [ranking for ranking in csv.DictReader(open('rankings.csv'))]

It will be convenient to have quick access to all the people ranked for a role, and all the roles a person is ranked for.

by_role, by_person = {}, {}

for ranking in all_rankings:
    by_role.setdefault(ranking['role'], []).append(ranking)
    by_person.setdefault(ranking['person'], []).append(ranking)

Now we can start using OR-Tools.

from ortools.sat.python import cp_model

model = cp_model.CpModel()
solver = cp_model.CpSolver()

The matching will be represented via a Boolean variable for each ranking that is true to select that match and false otherwise.

for ranking in all_rankings:
    ranking['selected'] = model.NewBoolVar('')

OR-Tools lets us use natural Python expressions to express constraints. First, we dictate that each role should have no more than the allowed number of people.

for role, rankings in by_role.items():
    total = sum(ranking['selected'] for ranking in rankings)
    model.Add(total <= spots[role])

Similarly, we require that each person is matched with at most one role.

for person, rankings in by_person.items():
    total = sum(ranking['selected'] for ranking in rankings)
    model.Add(total <= 1)

The objective will be to maximize the total score of selected matches.

total_score = 0
for ranking in all_rankings:
    score = float(ranking['score']) * ranking['selected']
    total_score += score

At this point, we need only ask for the solution.

status = solver.Solve(model)
print(solver.Value(total_score), solver.StatusName(status))
# 82.87644300144302 OPTIMAL

You can check that every role has its maximal number of people, and every person has one role (or zero; this example has more people than spots). The CP-SAT solver has done the heavy lifting of finding a solution and even guaranteeing that it is optimal.

with open('results.csv', 'w') as f:
    writer = csv.DictWriter(f, all_rankings[0].keys())
    for ranking in all_rankings:
        ranking['selected'] = solver.Value(ranking['selected'])

Code is also available in a notebook. Writing out results produces: results.csv

Role 23,Person 11,0.5238095238095238,0
Role 23,Person 41,0.5714285714285714,1

The approach here is general enough that with modifications it can be applied to a wide range of more or less related problems. There is always a question of whether the chosen metric is the right one to optimize, but making some choice can get you pretty far toward a good solution, with a solver that makes optimization easy. Indeed, what other problems might be formulated in a manner amenable to solving in this way? OR-Tools are handy!

Thanks to Maxime Labonne, whose Linear Programming course was very helpful as I started to use OR-Tools.