Rosenthal Atlanta

Embed Size (px)

Citation preview

  • 7/29/2019 Rosenthal Atlanta

    1/44

    1

    Ten Keys to Success in

    Optimization ModelingRichard E. Rosenthal

    Operations Research Department

    Naval Postgraduate School

    INFORMS Atlanta, October 2003

  • 7/29/2019 Rosenthal Atlanta

    2/44

    2

    Theme

    Optimization is valuable and pervasive (no need topreach to the converted here)

    Practical optimization applications continue to be:

    biggermore complex

    closer to real-time (if the situation warrants)

    less dependent on OR gurus

    more depended upon by companies, and

    taken for granted or

    taken over and claimed credit for by non-ORs

  • 7/29/2019 Rosenthal Atlanta

    3/44

    3

    Theme (continued)

    As we all know, this has been made possible byremarkable improvements in computers,modeling systems, and solvers (algorithms andtheir implementations). We have many greatresearchers and commercial implementers tothank.

    But, there is also another, important,

    sometimes overlooked piece of the story:GOOD MODELING PRACTICE

  • 7/29/2019 Rosenthal Atlanta

    4/44

    4

    Acknowledgements

    My ideas on good modeling practice have beengreatly influenced by my NPS colleaguesJerry Brown, Matt Carlyle, Rob Dell, KevinWood, and other great modelers I haveobserved in the practice of their art, such asHarlan Crowder, Terry Harrison, KarlaHoffman, David Ryan, Linus Schrage, JulieWard, Andres Weintraub, Kirk Yost.

    Many great ideas have come from NPSstudents.

    Todays talk owes a special debt to JerryBrown.

  • 7/29/2019 Rosenthal Atlanta

    5/44

    5

    Can You Teach Modeling?

  • 7/29/2019 Rosenthal Atlanta

    6/44

    6

    Key #1: Communicate Early and Often

    Mathematical formulation kept up to dateVerbal description of formulation

    Executive summary in the right language

    John Stuart Curry,Tragedy and Prelude - J ohn Brown

  • 7/29/2019 Rosenthal Atlanta

    7/44

    7

    Mathematical Formulation

    Index use

    Given data (and units)

    in lower case

    Decision Variables (and units)in UPPER CASE

    Objectives and constraints

  • 7/29/2019 Rosenthal Atlanta

    8/44

    8

    Verbal Description of Formulation

    Constraints [3] ensure that one service facilityis assigned responsibility for each product

    linep.

    You wonder why I mention this, but look at our

    applied literature.

  • 7/29/2019 Rosenthal Atlanta

    9/44

    9

    Non-mathematical Executive Summary

    Jerry Browns Five Essential Steps:

    What is the problem?

    Why is the problem important? How will the problem be solved without you?

    How will you solve the problem?

    How will the problem be solved with yourresults, but without you?

  • 7/29/2019 Rosenthal Atlanta

    10/44

    10

    Refining Your Executive Summary

    Have a non-OR read your summary out loud

    Ask the reader to explain what is going on

    Listen well

    Revise and repeat

    If you dont learn how to speak in theexecutives language, then someone lessqualified than you will be entrusted withsolving their problems.

  • 7/29/2019 Rosenthal Atlanta

    11/44

    11

    Non-mathematical Executive Summary

    Thats it? Thats peer review?

  • 7/29/2019 Rosenthal Atlanta

    12/44

    12

    Key #2: Bound all Decisions

    A trivial concept, too-often ignored

    Remember all the formal neighborhoodassumptions underlying your optimizationmethod?

    Bob Bixby tells of real customer MIP with only51 variables and 40 constraints that could notbe solved until bounds were added, and thenit solved in a flash.

  • 7/29/2019 Rosenthal Atlanta

    13/44

    13

    Bound all Decisions

    Optimization is an excellent way to find dataerrors, but it really exploits them

    Moderation is a virtue

    Bonus! You never have to deal with the

    embarrassment (or the theory) ofunbounded models.

  • 7/29/2019 Rosenthal Atlanta

    14/44

    14

    Bound all Decisions

  • 7/29/2019 Rosenthal Atlanta

    15/44

    15

    Key #3: Expect any Constraint toBecome an Objective, and Vice Versa

    Real-world models are notorious for multiple,conflicting objectives

    Expert guidance from senior leaders is ofteninterpreted as constraints

    These constraints are often infeasible

    Discovering what can be done changes yourconcept of what should be done

    Contrary to impression of textbooks, alternateoptima are the rule, not the exception.

  • 7/29/2019 Rosenthal Atlanta

    16/44

    16

    Key #4: Sensitivity Analysis in the RealWorld Is Nothing Like Textbook SA

    LP Sensitivity Analysis, Textbook Style

    Disappointing in practice because theory creates

    limits.

    Textbooks have sleek algorithms for one modificationat a time, all else held constant, e.g.,

    minimize SjcjXj+dXkNot very exciting in practice. So why is this stuff inall the textbooks? What is worth talking about?

  • 7/29/2019 Rosenthal Atlanta

    17/44

    17

    LP Sensitivity, Practitioner Style

    Operations Research, Jul-Aug 2002

  • 7/29/2019 Rosenthal Atlanta

    18/44

    18

    Sensitivity Analysis, Practitioner Style

    Large-scale LP for optimizing airlift -- multiple time-space muticommodity networks, linked together withnon-network constraints .

    Initial results on realistic scenario: only 65% of

    required cargo can be delivered on time.Analysis of result revealed most of the undeliveredcargo was destined for City A from City B, so.. what ifwe redirect some of this cargo to City A?

    Sensitivity analysis: add ~12,000 new rows and~10,000 columns... on-time delivery improves to 85%.

  • 7/29/2019 Rosenthal Atlanta

    19/44

    19

    Forget Textbook Sensitivity Analysis,Plan on Lots of Model Excursions

    The beauty of this is that it is only of theoreticalimportance, and there is no way it can be of any practicaluse whatsoever!

  • 7/29/2019 Rosenthal Atlanta

    20/44

    20

    Key #5: Bound the Dual Variables

    Huh?

    Elastic constraints,

    with a linear (or piece-wise linear) penalty per unit of

    violation,bound the dual variables

    Im willing to satisfy this restriction (constraint),

    as long as it doesnt get too expensive.Otherwise, forget it;

    Ill deal with the consequences

  • 7/29/2019 Rosenthal Atlanta

    21/44

    21

    Key #6: Model Robustly

    Your analysis should consider alternate futurescenarios, and render a single robust solution.

    There may be many contingency plans,

    but you only get one chance per year

    to ask for the money to get ready.

  • 7/29/2019 Rosenthal Atlanta

    22/44

    22

    Model Robustly

    This is the part I always hate.

  • 7/29/2019 Rosenthal Atlanta

    23/44

    23

    Key #7: Eliminate Lots of Variables

    Big models get to be big through Cartesianproducts of indices

    Find rules for eliminating lots of index tuplesbefore they are generated in the model

    Sources of rules: mathematical reasoning andcommon sense based on understanding of the

    problem

    You can often eliminate constraints too!

  • 7/29/2019 Rosenthal Atlanta

    24/44

    24

    Example 1 of Variable Elimination

    XD(a,i,r,t) = # of type a aircraft direct deliveringcargo for customer i on route r departing at time t

    Allow variable to exist only if

    Route r is a direct delivery route from customeris origin to is destinationAircraft type a is available at is origin at tAircraft type a can fly route rs critical leg

    Aircraft type a can carry some cargo type thatcustomer i demandsTime t is not before is available-to-load timeTime t is not after is required delivery date+ maxlate travel time

  • 7/29/2019 Rosenthal Atlanta

    25/44

    25

    Example 2 of Variable Elimination

    Ann Bixby and Brian Downs of AspenTechnology developed real-timeCapable-to-Promise model for large meatpacking company

    One of their major efforts to bring solutiontimes down low enough was variableelimination.

  • 7/29/2019 Rosenthal Atlanta

    26/44

    26

    Eliminate Lots of Variables

    You wont win the Nobel or Lanchester Prize

    for this key idea, but it really, really helps.

  • 7/29/2019 Rosenthal Atlanta

    27/44

    27

    Eliminate Lots of Variables

    In effect, what youre doing is taking a big lead offthird.

  • 7/29/2019 Rosenthal Atlanta

    28/44

    28

    Key #8: Incremental Implementation

    In a complex model, add featuresincrementally. Test each new feature onsmall instances and take no prisoners.

    When new features dont work, there iseither a bug to be fixed or a new insight to

    be gained. Either way, treasure the learningexperiences.

  • 7/29/2019 Rosenthal Atlanta

    29/44

    29

    Incremental Implementation

    Eliminate variables corresponding to airliftersswitching from long-haul to shuttle status, ifthere are no foreseeable shuttle opportunities.

    Feature tested with small example: removing theoption to make a seemingly foolish decision actuallycaused degradation of objective function.

    What happened?

    Euro

    US SWA FOB

  • 7/29/2019 Rosenthal Atlanta

    30/44

    30

    Key #9: Persistence

    Any prescriptive model that suggests a planand then, when used again, ignores its ownprior advice

    is bound to advise something needlesslydifferent, and lose the faith of itsbeneficiaries.

    Jerry Brown

  • 7/29/2019 Rosenthal Atlanta

    31/44

    31

    Illustration of Persistence

    Customer Eqpt Type Site Distance

    Cust01 Eqpt-01 HHH 353Cust02 Eqpt-01 HHH 724

    Cust03 Eqpt-01 HHH 773

    Cust04 Eqpt-01 YYY 707

    Cust05 Eqpt-01 YYY 719Cust06 Eqpt-03 RRR 495

    Cust07 Eqpt-01 HHH 442

    Cust08 Eqpt-03 RRR 590

    Results for Initial Case with 8 Customers

    There are initially 8 customers to serve. Wemust choose serving site and equipment type.

  • 7/29/2019 Rosenthal Atlanta

    32/44

    32

    Illustration of Persistence

    Just a moment after this solution isannounced, two high-priority customers callin. The model is rerun with the 10customers.

    There are not enough assets to cover all 10customers.

    The new solution requires a majorreallocation of assets. Major changes in thesolution are highlighted.

  • 7/29/2019 Rosenthal Atlanta

    33/44

    33

    Illustration of Persistence

    Customer Eqpt Type Site Dist (nm)

    Cust01 Eqpt-01 HHH 353

    Cust02 Eqpt-01 YYY 678

    Cust03 Eqpt-01 YYY 703Cust04

    Cust05 Eqpt-03 RRR 705

    Cust06 Eqpt-03 RRR 495

    Cust07 Eqpt-01 HHH 442Cust08

    Cust09 Eqpt-01 HHH 353

    Cust10 Eqpt-01 HHH 353

    esults for 10 Customers without Persistenc

    NOT SERVED

    NOT SERVED

  • 7/29/2019 Rosenthal Atlanta

    34/44

    34

    Illustration of Persistence

    A persistent version of the model isrun to obtain a new optimal solution

    that discourages major changes fromthe original announced solution.

    Add to objective function: penalties ondeviations from original solution,weighted by severity of disruption.

  • 7/29/2019 Rosenthal Atlanta

    35/44

    35

    Illustration of Persistence

    Customer Eqpt Type Site Dist (nm)

    Cust01 Eqpt-01 HHH 353

    Cust02 Eqpt-01 HHH 724

    Cust03 Eqpt-01 HHH 773Cust04

    Cust05 Eqpt-01 YYY 719

    Cust06 Eqpt-02 RRR 495

    Cust07 Eqpt-01 HHH 442Cust08

    Cust09 Eqpt-01 YYY 380

    Cust10 Eqpt-01 RRR 363

    Results for 10 Customers with Persistence

    NOT SERVED

    NOT SERVED

  • 7/29/2019 Rosenthal Atlanta

    36/44

    36

    Illustration of Persistence

    At a cost of 1% in the objective function, thepersistent solution causes no disruption to theannounced plans, other than substitution of thetwo new customers.

    Orig.

    With persistence? - No Yes

    Customers 8 10 10

    Objective function valueCustomers not served 0 180.00 180.00

    Original objective 206.67 151.75 155.20

    Total 206.67 331.75 335.20

    Comparison of SolutionsSubsequent

  • 7/29/2019 Rosenthal Atlanta

    37/44

    37

    Persistence

    Key reference:

    G. Brown, R. Dell, K. Wood, Optimization and

    Persistence, Interfaces1997.

  • 7/29/2019 Rosenthal Atlanta

    38/44

    38

    Key #10: Common Sense

    Heuristics are easy --- so easy we aretempted to use them in lieu of more formal

    methods

    Heuristics may offer a first choice to assess

    a common sense solution

    But, heuristics should not be your only choice

  • 7/29/2019 Rosenthal Atlanta

    39/44

    39

    Common Sense

    A formal optimization model takes longer todevelop, and solve

    But it provides a qualitativebound on eachheuristic solution

    Without this bound, our heuristic advice is ofcompletely unknown quality

    This quality guarantee is key

  • 7/29/2019 Rosenthal Atlanta

    40/44

    40

    Common Sense

    Its OK to use a heuristic,

    but you should pair it with a traditional,calibrating mathematical model

    With no quality assessment,

    you are betting your reputation

    that nobody else is luckier than you are

  • 7/29/2019 Rosenthal Atlanta

    41/44

    41

    Common Sense

    I tend to agree with you,

    especially since thats my lucky number.

  • 7/29/2019 Rosenthal Atlanta

    42/44

    42

    10 Keys to Success in Optimization Modeling

    #1 Write formulation,communicate with execs

    #2 Bound decisions

    #3 Objectives andconstraints exchangeroles (alt. optima likely)

    #4 Forget aboutsensitivity analysis as

    you learned it

    #5 Elasticize (boundduals)

    #6 Model robustly#7 Eliminate variables

    avoid generating themwhen you can

    #8 Incrementalimplementation

    #9 Model persistence

    #10 Bound heuristics withoptimization

  • 7/29/2019 Rosenthal Atlanta

    43/44

    43

    Thats it? Thats the Grand Unified Theory?

    10 Keys to Success in Optimization Modeling

  • 7/29/2019 Rosenthal Atlanta

    44/44

    44

    Questions and Comments?

    Stephen Hansen, Man on a Limb