INFORMS Open Forum

Expand all | Collapse all

AI and Prescriptive Analytics

  • 1.  AI and Prescriptive Analytics

    Posted 10-19-2018 15:42
    Should we include prescriptive analytics in AI?  Does the audience matter? 
    I've been trying to find an answer to this question. 

    Prescriptive analytics certainly provides intelligence for questions people have.  For quantifiable challenges, prescriptive analytics can do many things even better than what a person could do.  But as far as my reading shows, prescriptive analytics doesn't fit into the traditional AI definitions, which reflect machine learning and solutions to more subjective problems.  

    When I try to explain what I do to a general audience, however, AI seems to make more sense than prescriptive analytics.  AI, providing a way to automate tasks, is interesting and what they need.  Prescriptive analytics results in puzzled stares.

    Any suggestions or comments?

    ------------------------------
    Jane Mather, Ph.D.
    Principal, Critical Core
    Fraser CO
    ------------------------------


  • 2.  RE: AI and Prescriptive Analytics

    Posted 10-20-2018 08:34
    It is an interesting question.
    After working in this filed (confluence of New AI [a.k.a deep learning, reinforcement learning, and deep reinforcement learning], ML, Soft Computing, Evolutionary Computing, Fuzzy Computing, Applied Statistics, and Optimization) for 30 years, I come to the following conclusion:

    Prescriptive analytics, which aims at providing insights to the domain expert/decision maker, predominantly involves formulating and solving linear and nonlinear optimization problems. Thus, it is distinct from descriptive and predictive analytics. Quite easily one can relate predictive analytics (which comprises data/text/web mining techniques- that in turn involve ML. new AI and applied statistical techniques) to the catch-all word AI, used by you. 

    There is an umpteen number of use cases one can cite to explain the usefulness of prescriptive analytics without relating it to AI. 
    Therefore, AI and prescriptive analytics can co-exist without actually being related to each other. 

    Having said that Tabu search, a metaheuristic, is inspired by ideas from AI, while Hopfield neural network is designed to solve combinatorial optimization problems. These are the two links that come to my mind which relate these two fields.

    Best regards
    Ravi

    ------------------------------
    Dr. Ravi Vadlamani
    Professor and Head, Center of Excellence in Analytics,
    Inst for Dev. and Res.in Banking Tech (IDRBT), Hyderabad
    Hyderabad, India
    ------------------------------



  • 3.  RE: AI and Prescriptive Analytics

    Posted 10-20-2018 08:35
    Jane … AI and prescriptive analytics are not directly related. The continuum of analytics is: descriptive --> diagnostic --> predictive --> prescriptive. Prescriptive is associated with optimization. For example, for all of the "what if" scenarios from the predictive stage, the prescriptive stage answers which is the "best" choice. It may involve linear programming. AI is associated with robotic process automation (RPA), machine learning (ML), and cognitive software. They involve automating tasks that humans do. If you have 15 minutes watch the YouTube video "Humans need not apply".  … Gary … Gary Cokins

    ------------------------------
    Gary Cokins
    Founder and CEO
    Analytics-Based Performance Management LLC
    Cary NC
    ------------------------------



  • 4.  RE: AI and Prescriptive Analytics

    Posted 10-21-2018 17:56
    ​We have found AI methods with regards to proscriptive particularly helpful when a firm needs to respond to change or opportunity - what we call plan repair or restoring the health of the system.  Also some AI methods are helpful with matching assets with demand, especially closer to execution.  As Herbert Simon observed in his 1980's interface article - two heads can be better than one

    ------------------------------
    Kenneth Fordyce
    Director Analytics
    Arkieva
    Wilmington DE
    ------------------------------



  • 5.  RE: AI and Prescriptive Analytics

    Posted 10-20-2018 10:56
    I think you're right in suggesting the audience matters. As a consultant I am usually trying to distinguish between Machine Learning and AI. I find it useful to define AI as having a computer actually making a decision in place of a person. So, once you've handed the keys of the decision to a computer, that's what I call AI. 

    But I'm less of an AI / ML expert and more of an OR person. I see prescriptive analytics as (ideally) a process which lets you ensure that your AI is at least as good as the humans that have been making the decisions up to this point. But you can have analytics which prescribes action without automating it, and you can automate decisions without really doing the work to ensure it's a good idea.

    ------------------------------
    Zohar Strinka
    Denver, CO
    ------------------------------



  • 6.  RE: AI and Prescriptive Analytics

    Posted 10-22-2018 11:04
    I believe it is interesting to observe how other communities interpret what descriptive, predictive, and prescriptive analytics mean in the context of their disciplines. The following excerpt is from a blog post on Reinforcement Learning (RL) by Professor Ben Recht from Berkeley:

    "Descriptive
    analytics refers to summarizing data in a way to make it more interpretable. Unsupervised learning is a form of descriptive analytics. Predictive analytics aims to estimate outcomes from current data. Supervised learning is a kind of predictive analytics. Finally, prescriptive analytics guides actions to take in order to guarantee outcomes. RL as described here falls into this bucket."

    You can read the rest here: http://www.argmin.net/2018/01/29/taxonomy/

    ------------------------------
    Thiago Serra
    Visiting Research Scientist, Mitsubishi Electric Research Labs
    ------------------------------



  • 7.  RE: AI and Prescriptive Analytics

    Posted 10-23-2018 14:50
    Right, Thiago,

    "they" sometimes don't know about "our" interpretation, and a pessimistic estimate could be to find us outnumbered by "non-#orms" interpreters of prescriptive analytics and the #orms way will simply pushed aside, out of the way...

    :-/

    Marco


    ------------------------------
    Marco Luebbecke
    Professor of Operations Research
    RWTH Aachen University
    Aachen
    ------------------------------



  • 8.  RE: AI and Prescriptive Analytics

    Posted 10-24-2018 10:38
    Whenever I see "descriptive analytics" or "predictive analytics" or "prescriptive analytics" my train of thought briefly comes to a halt while I remind myself which one it is. I wish they had been named "data analytics" and "forecast analytics" and "decision analytics" which are shorter, more distinct, and more to the point.
    But maybe it's too late. In the leadoff talk at this month's INFORMS Regional Analytics Conference in Chicago, Mike Watson of Opex Analytics asserted that the use of "analytics" is already fading and that it's all called AI now. Perhaps soon I'll have to think of "data AI" and "forecast AI" and "decision AI".


    ------------------------------
    Bob Fourer, President
    AMPL Optimization Inc.
    Evanston, IL, USA
    ------------------------------



  • 9.  RE: AI and Prescriptive Analytics

    Posted 10-24-2018 02:24
    Yes, I agree with you. Reinforcement learning can also be considered a form of prescriptive analytics. It is actually optimizing one's actions on the go.

    ------------------------------
    Dr. Ravi Vadlamani
    Professor and Head, Center of Excellence in Analytics,
    Inst for Dev. and Res.in Banking Tech (IDRBT), Hyderabad
    Hyderabad, India
    ------------------------------



  • 10.  RE: AI and Prescriptive Analytics

    Posted 10-26-2018 03:17
    I may be a bit alone (and stubborn) with this, but it is my understanding of "optimization in practice" (as a mathematical optimizer by education and operations researcher by heart) that with optimization we often tackle "complex" decision situations.

    what do I mean by that? (ask also Ed Rothberg of Gurobi)

    we produce plans like timetables, detailed production sequences, warehouse locations, selections from huge sets, configurations with crazy constraints. the sheer amount of information such a plan gives to us (could be a plan for a year...) is enormous. finding such a plan would technically have meant to search a search space that is so ridiculously large no one could ever imagine a number to express the number of possibilities---combinatorial explosion. in optimization and operations research, or prescriptive analytics for that matter, we produce such a plan knowing that there is no better plan, or at least we know it's quality. (whatever "best" means, that is often debatable, but optimization offers help here as well). we give best decision support in super complex situations.

    machine learning also gives decision support (I exaggerate now to make my point). machine learning can forecast (and learn, and find patterns, and...) eg what the next word is that you type on your smart phone. or what the next move in Go should be. or whether it rains tomorrow or not. or what the probability is that a certain observation means cancer, or that a pattern in financial transactions are fraud, or that the sentence the AI just "read" actually translates into this and that sentence in Chinese. Let me say: these a SUPER SIMPLE decision situations. YES, ML/AI does provide me decision support, maybe even optimal. It gives me the most likely word as a suggestion that I will type next. Encoding this solution is super small. "easy"

    Let us not confuse the (current) application areas/decision situations. Both, ML/AI and Opt/ORMS have their merits. But we, as an OR community, are at risk that ML/AI steals "our brand" (which is decison support based on mathematical models and methods). We can embrace that and maybe we should. As we did with the analytics notion.  Then we may have best decision support for "simple" situations (like cancer diagnosis, this is "simple" because it is yes/no/probably) and for "complex" ones. I will keep reminding us that we have the tools for the complex decisions. We should not give this up easily.

    Therefore I cannot encourage us #orms people to call eg reinforcement learning "optimization". It somewhat waters down what optimization is capable of.









    ------------------------------
    Marco Luebbecke
    Professor of Operations Research
    RWTH Aachen University
    Aachen
    ------------------------------



  • 11.  RE: AI and Prescriptive Analytics

    Posted 10-27-2018 10:15
    Marco,

    I agree with what you said, but this raises another question: is "prescriptive analytics" more or less synonymous with "optimization", or is optimization a subset of "prescriptive analytics"? If an ML-based systems says "reject this transaction, it's probably fraudulent", it seems to be making a prescription, and it is based on analytics. As a fellow optimizer, it's not what I think of when I hear the phrase "prescriptive analytics", but to quote some old English dude, there are more things in heaven and earth than are dreamt of in my philosophy.

    ------------------------------
    Paul Rubin
    Professor Emeritus
    Michigan State University
    East Lansing MI
    ------------------------------



  • 12.  RE: AI and Prescriptive Analytics

    Posted 10-28-2018 01:18
    To the Analytics community:

    I don't want to throw improper cold water on the discussion of analytics,
    but for years I've wondered about the word itself.  It has always felt like a fancy
    buzzword that was cooked up by accident when "analysis" would have been fine.
    Can someone explain otherwise?

    ------------------------------
    Michael Saunders
    Stanford University
    Stanford CA
    ------------------------------



  • 13.  RE: AI and Prescriptive Analytics

    Posted 10-25-2018 01:41

    In a variety of topics, ML/AI has been also used to compute, prescribe and take decisions. It seems like with ML is also possible to find a way to learn how to make the right decisions and that ML is not just for predictive analytics. 

    Therefore, I think is a "yes", with ML/AI is also possible to do prescriptions. In the sense of: "something that is suggested as a way to do something or to make something happen"*. Which is related to the decision-making process.

    Here is an interesting article; I agree how the author describes the role of Optimization in ML/AI:

    https://www.ibm.com/developerworks/community/blogs/jfp/entry/Machine_Learning_As_Prescriptive_Analytics?lang=en

    *https://www.merriam-webster.com/dictionary/prescription



    ------------------------------
    Edgar Gutierrez-Franco
    Graduate Teaching Associate. PhD Candidate.
    University of Central Florida
    Orlando FL
    ------------------------------