For most of my career, I have been engaged in making decisions, or helping others make decisions. Way back in 1997, when I was just a boy, I used a financial tool known as the Hillier Model to help a large South African glass bottle manufacturer decide where to put capital in order to best serve market conditions. Naturally, I used a well-known spreadsheet software to crunch the numbers for me; it was the one that required a forward slash to bring up the menu. Think about that for a minute, old-timers.
This did a few things for me. Firstly, it taught me about uncertainty – not statistical uncertainty – human uncertainty. I worked day and night for three days running different scenarios based on different pieces of historical data. As the numbers started to point me in a direction, I got all ‘engineer’ on it. What about El Nino? What about South Africans turning away from beer to becoming wine connoisseurs? What if the historical data I was working with was rubbish? Of course I had done some regression to generate forecasts, some of which would be wrong, but some would be lucky. These weren’t numbers on a trend. Millions of dollars and hundreds of jobs were at stake.
Secondly I learned that the presentation of the information is often more powerful than the content. While the human is still casting the mark-one eyeball over certain decisions it is relatively easy to paint one set of results in a better light than the others, regardless of the ‘facts’. This is incredibly important, as some measures of human insight simply cannot be bottled or computerized.
Or can they?
We are all hearing about machine learning. It is the best thing since sliced cheese. It can tell you when you want sliced cheese. It can tell you who else is buying sliced cheese. And it does all of this without explicitly being programmed to do so, according to the propaganda.
As something of a decision nerd, I have watched the space of computerized decision support morph over the years. I remember getting excited about neural networks, before someone suggested they were just mathematical nuances of generalized linear models. A bit like realizing you didn’t need that $20000 BI tool after all, you could have done it with a Pivot Table.
So is machine learning on the same track? Will someone call out its mathematical sleight of hand and say ‘we’ve been doing that for years, we just called it Malcolm’? Our experience of machine learning as individuals tends to be when Amazon suggests you buy a box of unscented candles because you have just bought a new pair of Levis. Yes, it learned. What, I am not quite so sure, but it learned. And just how many people buy that pack of unscented candles? Refer to my point above on style over substance.
The other consideration is of course what constitutes machine learning in practical terms. I have been using SYSPRO’s Inventory Optimization for a number of years now and find that, within the statistical parameters set, with the right environment, it’ll predict what stock I should be carrying to achieve the service levels I need. Is that machine learning, or is that down to my skill as a practitioner in ‘programming’ the algorithms correctly? Strictly speaking it’s the latter – but for the guy on the street (or in the supply chain office), what practical difference does it make?
The thing is, when I was a boy trying to make it big in the corporates, I had a fledgling internet, sales numbers supplied on photocopies and a piece of software that disappeared into the ether 15 years ago or so. Now, everything is connected. The data is there. You don’t even need to own the software to make the learning happen. It will be wrong, but, on occasion, it will also be lucky.