The Boring Part of Bell Labs

It took me a long time to realize that Bell Labs was cool. You see, my dad worked at Bell Labs, and he has not done a single cool thing in his life except create me and bring a telescope to my third grade class. Nothing he was involved with could ever be cool, especially after the standard set by his grandfather who is allegedly on a patent for the television. It turns out I was partially right. The Bell Labs everyone talks about is the research division at Murray Hill. They’re the ones that invented transistors and solar cells. My dad was in the applied division at Holmdel, where he did things like design slide rulers so salesmen could estimate costs.[Fun fact: the old Holmdel site was used for the office scenes in Severance]But as I’ve gotten older I’ve gained an appreciation for the mundane, grinding work that supports moonshots, and Holmdel is the perfect example of doing so at scale. So I sat down with my dad to learn about what he did for Bell Labs and how the applied division operated. I expect the most interesting bit of this for other people is Bell Labs’ One Year On Campus program, in which they paid new-grad employees to earn a master’s degree on the topic of Bell’s choosing. I would have loved to do a full post on OYOC, but it’s barely mentioned online and my only sources are 3 participants with the same degree. If you were a manager who administered OYOC  or at least used it for a degree in something besides Operations Research, I’d love to talk to you (elizabeth@northseaanalytics.com).And now, the interviewElizabeth: How did you get started at Bell Labs?Craig: In 1970 I was about to graduate from Brown with a ScB in Applied Math. I had planned to go straight to graduate school, and been accepted, but I thought I might as well interview with Bell Labs when they came to campus.  That was when I first heard of the One Year On Campus program, where Bell Labs would send you to school on roughly 60% salary and pay for your tuition and books, to get a Masters degree.  Essentially, you got a generous fellowship and didn’t have to work as a teaching or research assistant, so it was a great deal.

I  got to go to Cornell where I already wanted to go, in the major I wanted, operations research.   Over 130 people signed up for the One Year On Campus program in 1970.  That was considerably more than Bell Labs had planned on; there was a mild recession and so more people accepted than they had planned.  They didn’t retract any job offers, but the next year’s One Year On Campus class was much smaller, so I was lucky.The last stage in applying was taking a physical at the local phone operating company.    Besides the usual checks, you had to look into a device that had two lighted eyepieces.  I looked in and recognized that I was seeing a cage in my left eye and a lion in my right eye.  But I also figured out this was a binocular vision test and I was supposed to combine the two images and see the lion in the cage, so that’s what I said I saw.   It’s unclear if Bell Labs cared about this, or this was the standard phone company test for someone who might be driving a phone truck and needed to judge distances.  Next time I went to an eye doctor, I asked about this; after some tests, he said I had functional but non-standard depth perception.    What did you do for Bell Labs?I worked in the Private Branch Exchange area.  Large and medium size companies would have small telephone exchanges that sat in their buildings.  It would be cheaper for them because most of the calls would be within the building rather than to the outside world, and thus avoid sending the call to a regular exchange a number of miles away and then back to the building.  You could also have special services, like direct lines to other company locations that you could rent and were cheaper than long distance charges.  The companies supplied their own phone operators and the operating companies were responsible for training, and the equipment and its maintenance and upgrades.  Most calls went through automatically e.g. if you knew the number.  But some would need an operator.  Naturally, the companies didn’t want to hire more operators than they needed to.  The operating company would do load measurements and, if the number of calls that needed an operator followed a Poisson distribution (so the inter-arrival times were exponential).The length of time an operator took to service the call followed an exponential distribution.

In theory, one could use queuing theory to get an analytical answer to how many operators you needed to provide to get reasonable service.  However, there was some feeling that real phone traffic had rare but lengthy tasks (the company’s president wanted the operator to call around a number of shops to find his wife so he could make plans for dinner (this is 1970)) that would be added on top of the regular Poisson/exponential traffic and these special calls might significantly degrade overall operator service.I turned this into my Master’s thesis. Using a simulation package called GPSS (General Purpose Simulation System, which I was pleasantly surprised to find still exists) I ran simulations for a number of phone lines and added different numbers of rare phone calls that called for considerable amounts of operator time. What we found was that occasional high-demand tasks did not disrupt the system and did not need to be planned for. Some projects I worked on: A slide rule for salesmen to estimate prices on site, instead of making clients wait until the salesman could talk to engineering.

Inventory control for parts for PBX.

I worked with a Ph. D.  mathematician on a complicated call processing problem.  I ran a computer simulation and he expanded the standard queuing theory models to cover some of the complexities of reality.  We compared results and they were reasonably similar.  Say more about inventory control?The newest models of PBX’s had circuit packs (an early version of circuit boards), so that if a unit failed, the technician could run diagnostics and just replace the defective circuit pack.  The problem was technicians didn’t want to get caught without a needed circuit pack, so each created their own off-the-books safety stocks of circuit packs.  The operating company hated this because the circuit packs were expensive, driving up inventory costs, and further, because circuit packs were being constantly updated, many off-the-book circuit packs were thrown out without ever having been used.  One operating  company proceeded with inspections, which one technician countered by moving his personal stock to his home garage. This was a classical inventory control problem, a subcategory of queuing theory.  I collected data on usage of circuit packs and time to restock, and came up with stocking levels and reorder points.  Happily, the usual assumptions worked out well.

After a while, the technicians were convinced they were unlikely to get caught short, the company was happy that they had to buy fewer circuit packs and they were accessible to all the technicians.  Everyone was happier.And the slide rule?While I was in graduate school, I became interested in piecewise linear regression (aka segmented regression), where at one or more points the regression line changes slope, jumps (changing its intercept) or both. I considered working on PLR for my Ph.D. dissertation.  On my summer job back, I saw a great fit with a project.  Salespeople would go out to prospective PBX customers but be unable to give them a quick and dirty cost estimate for a given number of phone lines, traffic load, etc.  It was complicated, because there were discontinuities:  for example, you could cover n phones with one control unit, so costs would go up linearly with each additional phone.  But if you had n +1, you had to have two control units and there would be a noticeable jump in costs.  There were a number of wrinkles like this. So the salesperson would have to go back to the office, have someone make detailed calculations and go back out to the customer, which would probably lead to more iterations  once they saw the cost.But this could be handled by piecewise regression.  The difficult problem in piecewise regression is figuring out where the regression line changes, but I knew where they were:  for the above example, the jump point was at n+1.  I did a number of piecewise regressions that captured the important costs  and put it on a ….I bet you thought I was going to say a programmable calculator.  Nope, this was 1975, and the first HP had only come out the year before.  I had never seen one and wouldn’t own one for two more years.  I’m not sure I could have gotten the formulae in the hundred line limit anyway. The idea of buying one for each salesperson and teaching them how to use them never came up.  I designed a cardboard slide rule for them.   I found piecewise regression useful in my work.  But that summer I recognized that research in the area had sputtered out after a couple of years, so I picked another topic for my dissertation.

Elizabeth: What did you do after your masters degree?Craig: I worked at Bell Labs for a year, and then started my PhD in statistics at UWMadison. There were no statistics classes worth taking over the summer, so I spent all four summers working at Bell Labs. How was Bell Labs organized, at a high level?I interviewed for a job at Murray Hill, where the research oriented Bell Labs work was done.  The job involved anti-ballistic missile defense and no secret details were gone into.  I didn’t get that job. I worked in a more applied location at Holmdel.I did go to one statistical conference at Murray Hill.  The head of the statistical area there was John Tukey, a very prominent statistician.  He simultaneously worked at Bell Labs and was head of the Princeton Statistics Department.  You don’t see much of that any more.There was a separate building in the area that did research in radio telescopes.  This was an outgrowth of research that investigated some odd radio interference with communication, that turned out to be astronomical.  I was never in that building.However, Bell Labs didn’t skimp on the facilities at Holmdel.  It had an excellent library with everything I needed in the way of statistics and applied math.  The computer facilities were also first-rate, comparable to that at the University of Wisconsin where I got my PhD.  Holmdel worked with the operating phone companies who provided actual phone service in their own geographical areas.  People at Holmdel would sometimes take exchange jobs at operating companies to better understand their problems.  One of these came back from a stint in New York City and gave a talk where he showed a slide of people pushing a refrigerator out of an upper story window of a derelict building while a New York Tel crew was working underneath them.A more common problem was that by the time I was there, technicians were not as bright as they had been.  A bright person who could not afford to go to college or maybe even finish high school in 1940 and had become a technician in an operating phone company had kids who could go to college, become engineers and be about to start work at Bell Labs in 1970.How was management structured?My recollection was that a first line-manager had a mean of 8 or 9 people.