INFORMS Open Forum

  • 1.  Machine Learning compared to computational probability and statistics from the 1970s and early 1980s

    Posted 03-23-2025 13:23

    As person from the "before times", I appreciate some elements of current machine learning.  Some of it strikes as material that has been known from the before times.  For example the terms one hot encoding versus Label Encoding.  These are essentially methods to make qualitative (or categorical) data into quantitative.  They are presented as new without recognition they have been around since the before times, trivial to do on programming environments such as APL (Array/Analytics programming language) .  AND without reference to burst of work done in the statistics community in categorical and ordinal categorical analysis - for example work by Agresti.  One hot encoding is a mainstay of binary formulations in MILP.  Assigning integers a mainstay of ordinal data analysis.

    Question - is something lost to the greater analytics community.



    ------------------------------
    kenneth fordyce
    director analytics without borders
    Arkieva
    Wilmington DE
    ------------------------------


  • 2.  RE: Machine Learning compared to computational probability and statistics from the 1970s and early 1980s

    Posted 03-24-2025 04:49

    You're very right, changing names is a sort of planned obsolescence



    ------------------------------
    Rosina Ramos
    lisboa
    ------------------------------



  • 3.  RE: Machine Learning compared to computational probability and statistics from the 1970s and early 1980s

    Posted 03-24-2025 12:00

    Good point, Ken.  Dynamic programming (now called Machine Learning or stochastic programming) has been around since the 60s when I took a course in it at Penn.  

    But the new ideas do add something new.  As in any mathematical discipline, we build on the work of earlier giants.



    ------------------------------
    Bruce Hartman
    Professor
    University of St. Francis
    Tucson, AZ United States
    bruce@ahartman.net
    website:http://drbrucehartman.net/brucewebsite/
    blog:http://supplychainandlogistics.org
    ------------------------------



  • 4.  RE: Machine Learning compared to computational probability and statistics from the 1970s and early 1980s

    Posted 03-26-2025 07:41
    Ken,
    Thanks for sharing your observation of forgotten history being reinvented and incorporated within the Machine Learning brand.
    In a related example, from Wikipedia....
    "The Hungarian method is a combinatorial optimization algorithm that solves the assignment problem in polynomial time and which anticipated later primal–dual methods. It was developed and published in 1955 by Harold Kuhn, who gave it the name "Hungarian method" because the algorithm was largely based on the earlier works of two Hungarian mathematicians, Dénes Kőnig and Jenő Egerváry.[1][2] However, in 2006 it was discovered that Carl Gustav Jacobi had solved the assignment problem in the 19th century, and the solution had been published posthumously in 1890 in Latin.[3]


    ------------------------------
    John Milne
    Clarkson University
    Potsdam, NY
    jmilne@clarkson.edu
    ------------------------------



  • 5.  RE: Machine Learning compared to computational probability and statistics from the 1970s and early 1980s

    Posted 04-01-2025 10:56
    Edited by Gerhard-Wilhelm Weber 04-01-2025 11:21

    Dear Dr. Fordyce,
    Dear Ken,
    Thank you very much again for this interesting, deep, and at the same time fresh impulse.
    Perhaps I can make a few additional comments. In fact, it always seems to me to be a good practice to view machine learning against the background of statistics,
    including optimal experimental design, and thus also mathematics in general. Machine learning was incorporated into
    statistical learning at a fairly early stage.
    I would like to mention the mathematical optimization theory foundations of machine learning via support vector machines.
    Furthermore, even today, mathematical modeling is often discussed when discussing statistics or machine learning.
    Statistical learning, like deep learning, interacted early on, even at the level of their motivations and inspirations,
    with brain research and other neurosciences, including cognitive science, so that the connection to artificial
    intelligence emerged early on.
    Over the years, some engineering disciplines have often referred to interfaces with all these fields as inverse problems. 

    These include more traditional tomography and discrete tomography, image generation and image processing, the same with
    video and sound, and, concerning all our senses, remote sensing, and the study of gravitational anomalies, which today
    leads us to earth sciences, space sciences, and cosmology. Early-Warning Systems and what I call Early-Chance Indicator Systems demonstrate that this does not exclude our OR-MS
    domain, but rather affects it and presents opportunities for it. These systems are used in all areas of our lives,
    including economics and business. Increasingly, hybrid systems, gene networks, metabolic networks, gene-environment networks, eco-finance networks, and
    regulatory systems, as well as stochastic calculus, are also helping here, and recently also with regime switching and
    paradigm shifting. As some of the above terms indicate, the entire field of machine learning and modeling is closely and mutually linked to
    optimal decision making. Not only machine learning, but OR-MS as a whole, including analytics, has enormous potential to make a unique contribution
    and serve sustainably in the future. Many thanks to you again, and to the whole INFORMS team, which faithfully provides this excellent forum for fruitful exchange.
    With kind regards, best wishes, Willi © Gerhard-Wilhelm Weber, "Times and Lives" (in completion)



    ------------------------------
    Gerhard-Wilhelm Weber
    Professor
    Poznan University of Technology
    Poznan
    ------------------------------