#34 Geoffrey De Smet: Solving Real-Time Scheduling Problems

So, um, first things first is that whenever you create a plan the day before, it's never really going to go in execution the way you planned it, right? Yes.

Um, but I like, there's a quote around Churchill that I like, and he said, like, "Plans are useless, but planning- Yes ... is essential." And it really brings down to the fact that if you don't have a plan, it's going to become chaos.

But if you have a plan where you try to be as accurate as possible, you're going to have a good starting plan, and then during the day it will change.

So one of the things we support is we first to make that go for that perfect optimal plan, right? That's when we solve for, you know, 30 minutes if you have a very big data set and things like that.

Yes.

Um, but then during the day, um, you're basically in survival mode.

Um, somebody's trapped in an elevator and one of your elevator technicians needs to drop everything and get those persons out of that elevator, right? Uh, there's no debate around that.

But, but you then need to figure out who's going to do that, right? So you need an answer immediately.

And that's where we can do, uh, we, we run warm start.

So basically, we take the solution we had that's being executed, and we, uh, calculate, okay, if we make this problem change, we now need to send somebody there immediately.

Who is the best person to send, and how will this affect the rest of the plan? And this is basically what we do during the day, right? And that's how you can change how the real- uh, the reality as it shifts and as it changes from the original problem- Yeah ...

statements during the day, right? Um, now, taking into account uncertainty is the next step.

Now, there's a big problem with taking into account uncertainty.

The first problem is that it's, that most, uh, companies do not have, like, distributions on their un- uh, the, uh, uncertainty.

So for example, um, if you say it will take on average, um, two hours to do that job, right? Very few companies will actually be able to tell you, like, okay, um, but the, it- it's not a bell curve, it's more like towards...

It, it takes on average like an hour and a half, but some cases take eight hours, right? So without that data- Right ...

you, you're dead in the water.

There's no point in, like, doing stochastic optimization and things like that.

Um, when you start taking this into account, you can of course make sure that, like, we, a couple of constraints we do to, to deal with uncertainty is, in field service routing, is, uh, we, the last visits are never priority visits.

So if you say, for example, "I need to do this visit today," we will actually make sure this happens in the early part of a technician's route because the, um, the, the early visits are always done.

It's always the last visit that's like more like a 50/50 case whether there will be enough time to do it or not, right? So you...

That's how you put in these kinds of r- uh, requirements to make sure that you skew it towards a case when, when uncertainty hits, when things don't happen, that it actually works out fine, right? Um, now, on our solver level, we can actually, we, we can actually deal with deeper uncertainty when you have these kinds of curves.

We, we can, we can have a lot of freedom where we can start adding these kind of things in there.

But I'm not convinced it's a practical thing to do today.

I've, I've not seen it in production, and, uh, simply because the data's there, is not there, the, the companies do not have the data in most cases.