Last week I was pretty optimistic about converting my big scheduling project’s constraints over to add_automaton. The approach was fairly clean, has a built-in way to generate visual aids, I could see maybe how to build a UI to use a JS graphing library to let the user build constraints, and I had proven to myself that I could decompose a big set of constraints into smaller, independent add_automaton calls. But then I hit a snag—the running time.

After the success of my experiment at splitting up the DFA documented in my last post, I now want to try more difficult constraints. Specifically, I was wondering about how to implement a constraint that gives everybody at least one two-day “weekend” every two weeks or so.

Building upon my last post, I am still talking about rostering, but now I want to see about using it in a project that I am working on. The problem I see with the nurse rostering example from my previous post is that it is a very complicated graph. What I want to do is to break up the graph into smaller, atomic parts, and see if it will still work. In fact it does work, and this post describes what I did.

I have a long-term client who needs to schedule doctors according to their capabilities, but also according to a complex set of rules. I developed a working solver years ago using OR-Tools and CP-SAT. Everytime they ask for some small improvement, I take the opportunity to revisit my core abstractions to see if I can’t figure out a more elegant way to approach the problem. Recently I’ve been studying a published example of nurse rostering that uses the add_automaton() function from CP-SAT. I believe this is a good approach that will fix all the things, so I am writing this series of blog posts. In the end, I think I’ve unvented something relating to the DFA graph’s dual, but we’ll see if that has legs.

Recently I have been working on better debugging of CP-SAT models. In order to test the various functions, I wanted to load a saved model and find variables and constraints. It turns out, I haven’t ever loaded a model from a pbtxt dump file. I searched in the forums and GitHub issues, and the best advice I found didn’t work. This short blog (Hello 2026!) is about how to load a CP-SAT model from a protobuf text dump in Python, as of version 9.16 of OR-Tools.

This post is mostly to try to get myself back in the habit of posting new blog content. I was looking over some of the examples in the OR-Tools repository, and I started playing around with the tasks and workers assignment problem. The original example can be found here: https://github.com/google/or-tools/blob/master/examples/python/tasks_and_workers_assignment_sat.py. In this post I go through the code bit by bit to understand what is going on, and then try some small changes to see what happens. I hope a newcomer to OR-Tools or CP-SAT might find this useful for starting their own exploration of the examples in the repository.

I was working through an example CP-SAT program on sports team scheduling, and I found a bit of AddBoolOr logic that I didn’t understand. So I wrote up a short program to explore what was going on.