Machine scheduling, operations research, production, and machine learning. An anecdote thread for @sidwindc.
Biographical background: from age 15 to about 21, I spent my summers working for the laboratory equipment manufacturer that also employed my dad (in "construction").
Biographical background: from age 15 to about 21, I spent my summers working for the laboratory equipment manufacturer that also employed my dad (in "construction").
After that I moved to prepoduction at Siemens, also for healthcare equipment, mostly coding production and tooling processes in CNC. This is where I stayed until my master's thesis.
I might be somewhere in this picture.
I might be somewhere in this picture.
For my master's thesis, and this is the topic of this thread, I moved to what is now Novartis in Basel into the scheduling group. This was both moving from a blue collar to a PhD level white collar environment and from the Bundesliga to the Champions League.
A chemical and pharmaceutical plant in a densely populated area along the Rhine river, and there are quite a few of them, is a zero tolerance environment. If something goes wrong, people die. If people die, scheduling carries the brunt of the blame.
So what is operations research? A bunch of mathematical planning tools, mostly various forms of optimization algorithms, to design a set of industrial processes. Typically routing, queueing, matching, assignment, transportation, flow, scheduling, supply chain inbound & outbound.
Machine scheduling, aka job shop or flow job scheduling, is the crown of operations research, and it is roughly the combination of all the things I mentioned. How to route multiple multi-stage batch processes through a complex machine park. This is the lifeblood of production.
But it's not only production, public transportation scheduling like train routing also depends on machine scheduling. Semiconductor design, which is really "a factory on a chip" does this on an insane zero-tolerance level.
Since the problem is so complex, crucial and expensive, companies will throw the kitchen sink at it. Every new technology will get a look. Even blockchain.
My job for the thesis was to look at Neural Networks, then in its Rumelhart-McClelland inspired second hype cycle.
My job for the thesis was to look at Neural Networks, then in its Rumelhart-McClelland inspired second hype cycle.
The task in a chemical plant is to map the many multi-stage chemical processes to the many chemical processors, all in exactly the right order, all with setup requirements, to maximize thruput, without anything exploding, ever.
My job was to take a somewhat simplified problem (we cut out a few real-world complexities to make it tractable) and see if it could be translated into a dynamic pattern recognition task. If there is a hole of a certain shape in the current schedule, which process fits there?
My answer was somewhat measured endorsement, and I believe machine scheduling is still a major driver in quantum computing research, since most problems in machine scheduling are combinatorial: one thing wrong means everything is wrong.
So this is actual production. Supply chain is production but geographically dispersed (routing becomes important) and vertically disintegrated (contracting becomes important), plus inventory planning plus distribution. Economic production is basically a carton version of this.
At the core of machine scheduling is network flow, for which we used graph theory, and queueing & flow theory, for which we used Markov chains and similar. So when I moved to the US with this background I had about 20 years of knowledge advantage over the economists at the time.
So while ostensibly I switched to economics I actually did operations research applied to economic problems, like Ken Arrow, Lloyd Shapley, David Gale, Tjalling Koopmans, and many folks at RAND and Cowles before me. Hence...
After graduating from Berkeley I moved back into transportation research (the "new mobility") and didn't get back into supply chain until a couple of years ago, when we built the Forrester tool.
This thread was brought to you by the #BeigeAcademy of unearthing forgotten disciplines that helped shape the post-War success of the US industrial base (and post-War economics).
@threadreaderapp unroll plz!
@threadreaderapp unroll plz!
Bonus anecdote: Since I came to this from a Herb Simon logic-based AI background and was somewhat suspicious of NNs at first, I initially proposed a simple solution in PROLOG. My adviser wiped it off the table with a simple four-part process. This is how I was won over to NNs.
Bonus bonus anecdote: At Siemens we had a behemoth CNC machine where three milling positions were connected thru the same drive shaft and the same tool feeder. Every time one of those things went down the whole behemoth went down. That's combinatorial.
