A Basic Branch and Bound Solver in Python using Cvxpy

Branch and bound is a useful problem solving technique. The idea is, if you have a minimization problem you want to solve, maybe there is a way to relax the constraints to an easier problem. If so, the solution of the easier problem is a lower bound on the possible solution of the hard problem. If the solution of the easier problem just so happens to also obey the more constrained hard problem, then it must also be the solution to the hard problem. You can also use the lower bound coming from a relaxed problem to prune your search tree for the hard problem. If even the relaxed problem doesn’t beat the current best found, don’t bother going down that branch.

A standard place this paradigm occurs is in mixed integer programming. The relaxation of a binary constraint (either 0 or 1) can be all the values in between (any number between 0 and 1). If this relaxed problem can be expressed in a form amenable to a solver like a linear programming solver, you can use that to power the branch and bound search, also using returned solutions for possible heuristics.

I built a basic version of this that uses cvxpy as the relaxed problem solver. Cvxpy already has much much faster mixed integer solvers baked in (which is useful to make sure mine is returning correct results), but it was an interesting exercise. The real reason I’m toying around is I kind of want the ability to add custom branching heuristics or inspect and maintain the branch and bound search tree, which you’d need to get into the more complicated guts of the solvers bound to cvxpy to get at. Julia might be a better choice.

A somewhat similar (and better) project is https://github.com/oxfordcontrol/miosqp which doesn’t use cvxpy explicitly, but does have the branch and bound control in the python layer of the solver. There are also other projects that can use fairly arbitrary solvers like Bonmin

As a toy problem I’m using a knapsack problem where we have objects of different sizes and different values. We want to maximize the value while keeping the total size under the capacity of the bag. This can be phrased linearly like so: \max v \cdot x s.t. \sum_i s_i x_i<= capacity , x \in {0,1}. The basic heuristic I’m using is to branch on variables that are either 0 or 1 in even the relaxed solution. The alternative branch hopefully gets pruned fast.

This is at least solving the problem fairly quickly. It needs better heuristics and to be sped up, which is possible in lots of ways. I was not trying to avoid all performance optimizations. It takes maybe 5 seconds, whereas the cvxpy solver is almost instantaneous.

Leave a Reply

Your email address will not be published. Required fields are marked *