[SCIP] About variable fixing during branching

[Hoen, Alexander]

Hello Fabian,

You might find this paper helpful: https://arxiv.org/abs/2412.14710, or even more so the accompanying code: https://github.com/alexhoen/bnbanalyzer.
Although it’s written in C++, it should still serve as a useful reference.

In that project, I collect all leaves of the Branch-and-Bound tree using the event handler system and then perform analyses on them.
If you remove the analysis part, the code will effectively print the entire Branch-and-Bound tree. You can either reuse the existing code directly or adapt it as a debugging tool for your Python implementation.

My suggestion is to ensure that you have parameterized it correctly as Branch-and-Bound (B&B). The relevant parameterization can be found in the SCIPInterface.

Hope this helps!

Best,
Alexander



________________________________
From: Scip <scip-bounces at zib.de> on behalf of Badilla Mera, Fabian <fbadilla at gatech.edu>
Sent: Tuesday, October 14, 2025 4:45:31 PM
To: scip at zib.de
Subject: [SCIP] About variable fixing during branching

Hi Everyone,

I'm currently using PySCIPOpt to generate B&B trees, because I'm interested in some information related to the leaves of the tree. I'm keeping track of the solving process using an Event handler that saves data such as the branching performed at each node, the node's LP objval, the dual solutions, among other information.

I want to use the most basic B&B algorithm possible to generate the full tree. To do this, I'm currently deactivating some functionality from SCIP using the methods: setPresolve(off), setHeuristics(off), setSeparting(off) and disablePropagation(). Also, I'm setting additional parameters like propagating/maxrounds, propagating/maxcuts to 0 and using DFS node selection.

My problem is that I've encountered some intances where the LP that SCIP is solving at the node is not the one that I expected, because it has additional variables being fixed. I think this happens whenever we found a solution in a Branch of the tree, and then SCIP use that information to fix variables on the rest of the tree. This leads SCIP to prune the branch before finding a node where the LP bound is worst than the global bound, because it marks the node as infeasible (even when the LP it is feasible).

In summary, I'd really appreaciate if you could guide me on which parameters I have to tweak to be able to get the full trees and deactivate the additional variable fixing. Also, if someone could explain me what conditions triggers the event "NODEINFEASIBLE" that would help me a lot, because in some cases I get that event even when the node's LP is feasible. How does SCIP detects that the branch will be infeasible for the original MIP?

Best regards,
Fabian Badilla
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20251014/29d74c7b/attachment.html>