Working in healthcare, labor scheduling is a critical part of data analysis. It’s important to meet demand while not scheduling unneccessary shifts.
When framing these problems, I look for inspiration to the grand-daddy of logistics: UPS. They’re close to solving one of the toughest problems in all of operations: the travelling salesman problem.
Your GPS may make it look easy, but finding the quickest route between a handful of destinations is extremely difficult. UPS is devoted to solving it through Project Orion.
I’ll never ascend to the heights of UPS genius, but I can try! Here are some data modeling takeaways from Project Orion.
Don’t rely on personal anecdotes…
There are millions of routes to choose from in the travelling salesman problem. Think you can evaluate every alternative?
We can only account for about seven numbers at a time. So juggling millions of options is a challenge.
Personal anecdote and mental math isn’t enough. Computational models let us compare many options without bias.
But don’t rely on the algorithm, either.
“The model made me do it!”
At the same time, you don’t want to drive the van into a lake.
Sometimes, things truly are just done a certain way just out of dumb habit. Other times, though, there’s good reason.
Don’t blindly follow the model’s outputs. If something seems off, check your assumptions. Is there an unreasonable relationship beween variables? Unrealistically high sensitivity? Remember who’s driving the car here — you! “The model made me do it” doesn’t cut it.
In other words, aim for the modelling “Sweet Spot.”
Data analytics’ “zone of impact” lies between what we already know and what we don’t believe.
Project Orion operates within this zone. The algorithm may come up with some counterintuitive routes. The driver is under no obligation to follow these routes — but he will be asked to justify his decision.
This system allows Orion’s insights to fall within the zone of impact.
Optimize AND Simulate
One temptation is to build a model without any accounting for variation. Sure, you’ve got a staffing model that will perfectly accomodate demand on an average day.
But what are you going to do on Christmas? Or the week after Christmas? Surely these two weeks will not have the same demand — so you should not assign the same workload.
I took a course entitled Simulation & Optimization. I wasn’t quite sure why these topics went together. Now it really makes sense. Optimizing for the average day really means optimizing for no day — because no day is average.
Your Changes Make Changes.
This is what makes makes the social science (and yes, this is economics — which is a social science) different than the physical sciences. Ultimately, we are looking at human action, which doesn’t always square with tidy models.
For instance, say you concluded that to optimize staffing, you’d like to move the starting shift time by a couple of hours.
That change in itself will trigger other changes. There could be behavioral changes. The staff may react poorly to a new start time, or and some people may decide to leave. This will impact your staffing.
This isn’t thermodynamics … there are few static laws in the social sciences.
Feasible Solution <> Implementable Solution.
This is the single most important takeaway. Sometimes, you just can’t execute the best solution. Maybe your model suggests split or reduced shifts. That could reduce the staffing pool, but is it great for employee morale? Can you really schedule a shift for a person-and-a-half?
Rather than tear up the model and start from scratch, incorporate the best of what’s possible.
This exercise of a continuously evolving, not-quite-optimal algorithm is called heuristics. From a WSJ piece on Orion.
“Instead of searching for the optimal, or best possible answer, heuristics is the search for the best answer one can find, the results continually refined over time, based on experience.”
“Look through, not at!”
Heuristics makes me think of a not-quite-fashionable pair of glasses. Sure, they may be ugly to look at. But what’s it like to look through them? That’s the real power of the model — how well it functions, not how perfect it looks.