/  15
Ann Oper Res (2007) 149:195–208
DOI 10.1007/s10479-006-0095-7
IP over 40+ Years at IBM Scientific Centers
and Marketing
Kurt Spielberg
Published online: 10 January 2007
C
Springer Science+ Business Media, LLC 2007

Time passes indeed! My days at IBM reach back to the late fifties. I was diverted from a career as Physics Professor at CCNY to one at IBM by the attraction of early computing. We worked in several locations not far from IBM head quarters at 590 Madison Avenue. Eventually, IBM formed a “mathematics and applications” group. It did advanced work in many areas, for example there were projects on “reading texts being converted to Braille output (Anne Schack), fabrics being produced on wool looms from patterns fed into computers (Janice Lourie), also “Autopromt”, an automatic machine tool generation project (Sam Matsa), and other such applications never done before on computers.

Those were heady days: whatever one did was new and exciting. I wrote some of the first elementary function subroutines of FORTRAN in assembly language, using Chebyshev functions for polynomial approximations, then rational expressions and continued fractions, to assure rapid and accurate computation. Corrections were in binary form, and one might have to fly to Cambridge to get available computer time at MIT! Eventually I headed a small team for developing a first full Package for FORTRAN.

It was good to work with the Russian mathematician Dr. Kogbetliantz, a colleague of Hadamard, who was quoted in the yellow Springer Math. Book, Vol. 1 (“Infinite Series”, Knopp). He was a kind old man (probably my age now!) who taught me much about the approximation of functions. We worked in a room in New York City also used by the team who had recently designed and developed FORTRAN, a major task. I became friendly with members of that illustrious group, such as Don Quarles and Harlan Herrick.

Later, a first major project was on Global Weather Forecasting, jointly with the U.S.Weather Bureau. We used partial differential equations (suggested to the Weather Bureau first by John von Neumann). I eventually derived them from a general tensor form for a rectangular coordinate system (a paper on it was published in the IBM Research Journal; I believe only a few US Weather Bureau members ever read it).

We used a grid with ten-thousand points in a plane and nine vertical layers to represent
the Northern Hemisphere. Ingredients of weather (temperature, humidity, wind components)

K. Spielberg ()
University of PA, Wharton School, Philadelphia, PA 19104, USA
e-mail: kurtspie@yahoo.com

Springer
196
Ann Oper Res (2007) 149:195–208

were propagated via difference equations; energy variations from long and short wave radi- ation and turbulent diffusion were calculated to represent changes within the quads of the grid points.

The “Mathematics and Applications Department” became the NY Scientific Center (one of 5 U.S. Centers; there were also 4 or 5 international ones). We gave three-week classes in London, Paris, Tokyo, weekly topics, such as...., were typically presented by half day lectures. I exchanged a first class round trip ticket to Tokyo (to give lectures with simultaneous translation at Tokyo University) into a round-the-world trip, Hongkong, Bangkok, Tel Aviv, etc., a great experience! In Paris I lectured on Navier-Stokes equations, solved as difference equations on the “Stretch” Computer, to top European partial differential equation specialists, who probably came also (or mainly) for “April-in-Paris”.

Switching to IP

We then looked for a new project: our library provided an article by J.F. Benders (1962), on “Benders Partitioning” inNumerische Mathematik (where I had a paper on approximations). Could it be applied, say, to integer programming? We became aware of novel work by Gomory on cuts, and of attempts to solve the TSP problem. I found many ways to specialize the Benders scheme, proving several formulations of important Integer Programming problems, some of which I found later in the literature,

When work on the TSP yielded little, I tackled the transportation problem (TP) with fixed charges on nodes. I had earlier coded a dual solver of the TP with results which purported to be better than obtained by a primal version a friend had coded (they were quoted in the famous Ford-Fulkerson book). I wrote a paper for an ACM conference, giving examples solved by a simple code with a sequence of TP’s. But the problem did not seem to be tractable for practical sizes (and I could not find anyone with real examples).

The easiest wayto chicken out was a switch to a problem with fixed charges onlyat sources, i.e., “uncapacitatedplant location”, whichI recklessly named “simple plant location problem” (SPLP). For SPLP I developed enumerative codes governed by a “greedy” approach (Spiel- berg, 1969a) with plants initially open; a forward branch would be a closing which produced maximal gain (certain “gain functions”, coefficients of “Benders inequalities”, could be up- dated highly efficientlyat each node). This algorithm was later used heuristically in the Lanch- ester Prize paper of Cornuejols, Fisher, and Nemhauser (1977). Gain functions also played a role in maximizing submodular functions, e.g., Fisher, Nemhauser, and Wolsey (1978).

A second code started at a “generalized search origin” with some plants open and others closed. It showed that a “good” initial node might require a much reduced number of nodes (1969b). I treated the problems (simple or capacitated) in a “disaggregated” (“DA”) form, i.e., with “DA” constraintsxijM.yirather than jxijb j.Michel Balinski had a “DA” form for the Plant Location problems already in earlier papers which we only saw later (Balinski, 1961, 1965). As first editor ofMathematical Programming he asked me to be an associate editor. We co-authored a survey of integer programming methods in Balinski and Spielberg (1969).

Cooperations, Carlton Lemke, enumeration for MIP, logical processing (probing)

The NY Scientific Center activities in Math. Programming, in research and applied work, were respectable even when compared to activities at IBM Research (Yorktown Heights) of leaders such as Gomory, Gilmore, Hoffman, Karp and Johnson. Mike Held worked with Dick Karp on Lagrangean methods, Mike Grigoriadis, John Greenstadt and Yon Bard worked in

Springer
Ann Oper Res (2007) 149:195–208
197
NLP; Susan Hahn and Paul Gilmore treated cutting stock problems; Bill White and Abel
Bomberault tackled important applications, e.g. railroad scheduling.

There were many distinguished visitors, such as Egon Balas, Monique Guignard, Harvey Salkin, later Laureano Escudero; and interactions with other centers (Berkeley, Los Angeles, Paris, Heidelberg, etc.). Monique and I applied ideas from plant location work to other problems, (“weakly linked”) with “DA” constraints (Guignard and Spielberg, 1968).

We wrote enumeration codes for (0, 1) IP, plant location and alike, and simple (branch and bound), “BB”, codes, e.g., “BB-MIP”, using Land and Doig ideas (1960). Dick Shareshian and I submitted FORTRAN codes to a Share library for wide dissemination, with write ups in scientific center reports. A survey of enumeration was in the proceedings of a “DO77” symposium in Vancouver-Banff (1979).

Carlton Lemke (of dual simplex method, complementary pivoting, etc., fame) was a guest from 1967 to 1969. It was a great pleasure to work with such a profound thinker. He was one of the kindest men I ever met (even though a paratrooper as young man, as one would not have guessed); we became life-long friends, but were much different collaborators: he a profound thinker and mathematician, I mainly concerned with practical ends. He wanted me to collaborate with him on complementary pivoting work and I still regret that I did not take him up on it. We kept in touch for many years, but, sadly, he started to suffer from Alzheimer’s disease and died in April 2004.

We extended an enumeration method of Balas (1965) to (0, 1)-MIP, including elements of Benders partitioning (Lemke and Spielberg, 1969). It seemed better than direct partitioning. Some of our tools were later used in conjunction with BB. I always placed emphasis on inequalities to retain the main advantages of Benders partitioning.

At a nodet we could fix some variables. Given an objective functionz=j(c j.y j+···), withyjbinary variables andz a lower bound onz, fixy jat 0 ifzt+c jz(“ceiling test”); i.e., one would have obtained a “degree 1” inequalityyj0 (withy jcomplemented variables this would permit the possible fixing of an original variable at 1). Also, one could look ahead for several branch choices to potential next nodes, sayt+ 1, via “double ceiling tests”. I.e., settingyjto 1 att and thenykto 1 att+ 1 could yield a degree 2 implication:y j+yk1, etc. With variables complemented one could get all four possible degree 2 covers. This showed it might make sense to generate, at any node, a collection of “minimal covers”, i.e. inequalities over (0, 1) variables: a set of low “degree” covers, which could provide important “logical information”. It was the start of “logical processing”, elaborated later at the Philadelphia Scientific Center with the central idea of “probing”; for a summary see Guignard and Spielberg (1981).

IBM Germany organized regular technical meetings, in the Black Forest (Wildbad). At these I would talk about IP and Logical Processing, I remember a fine talk by Arthur Geoffrion on Lagrangean Relaxation (he was also a good piano player). One positive effect was that I came to know the German HQ technical management well and was able to invite young German scholars to come to our centers as guests, financed by IBM Germany.

With the Paris Scientific Center we established regular interchanges of visitors, beneting both sides. I remember vividly three months in Paris in 1968, which were interrupted by the 1968 “student revolution”. Work came to a halt, as most traffic and garbage collection ceased for two months. I may have been the only person at IBM to work, walking from Passy (near the Eiffel Tower) to IBM near City Hall, 8–10 miles one way, past overturned cars and the Bourse set on fire.

The revolt ended with a spectacular parade up the Champs Elysees in support of General
De Gaulle, led by Georges Pompidou, which I attended with pleasure. The contacts in France
Springer

Share & Embed

More from this user

Add a Comment

Characters: ...