1 / 40

Partial Capture Location Problems: Facility Location and Design

Partial Capture Location Problems: Facility Location and Design . Dmitry Krass Rotman School of Management, University of Toronto With Robert Aboolian , CSUSM Oded Berman , Rotman. Rotman School of Management Ph.D. Program in Operations Management. Rotman

lawson
Télécharger la présentation

Partial Capture Location Problems: Facility Location and Design

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Partial Capture Location Problems:Facility Location and Design Dmitry Krass Rotman School of Management, University of Toronto With Robert Aboolian, CSUSM Oded Berman, Rotman

  2. Rotman School of ManagementPh.D. Program in Operations Management • Rotman • MBA program is among top 50 world-wide (FT, 2013) • #11 International MBA programs (BW, 2012) • Ranked #8 in research (FT, 2013) • Ranked #9 Ph.D. program among all business schools (FT, 2013) • University of Toronto • Ranked #1 in Canada • Ranked #16 in the world by reputation (Times of London, 2012) • RotmanPh.D. Program in OM • 1-2 students per year (25-50 applicants) • All students fully funded ($26K per year) for up to 5 years • Research areas: Supply Chain Management, Queuing, Revenue Management, Location Theory, other OR/OM topics • Our goal is 100% academic placements

  3. Rotman Ph.D. in OM If you have great math skills, are interested in applying them to important managerial problems, want to participate in world-class research, and are interested in a career as a university professor, Contact Dmitry Krass krass@rotman.utoronto.ca program web site: http://www.rotman.utoronto.ca/Degrees/PhD/Academics/MajorAreasofStudy/OperationsManagement.aspx

  4. Overview • Introduction • Facility Location Problems – a quick review • Partial Capture Models: a Unifying framework • Modeling design aspects • Single-Facility Design Problem • Non-linear knapsack approach • Sensitivity analysis • Multi-facility Design and Location Problem • Tangent Line Approximation (TLA) approach to non-linear knapsack-type problems • Iterated TLA method • Conclusions and Future Research

  5. Location Models – Brief Overview • Key interaction: customers and facilities • Application areas • Physical facilities: public, private • Strategic planning • Marketing (perception space), communications (servers, nodes), statistics/data mining (clustering), etc.

  6. Competitive Location Models: basics • Facilities always “compete” for customer demand • “Competitive location models” assume (at a minimum) • Customer choice (not directed) assignments • Not all facilities controlled by the same decision-maker • Goal is to maximize “profit” for a subset of facilities • Facilities outside the subset belong to “competitors”

  7. Modeling Competition • Static models • No reaction from competitor(s); “follower’s model” • “Dynamic” models (“stackelberg games”) • Some form of competitive reaction • Leader’s problem; Leader/follower/leader, etc. • Nash games (simultaneous moves) • Issues: non-existence of equilibria, solution difficulty, limited insights ✓

  8. Location Theory: Key literature • M. Daskin, 1995, “Netowork and Discrete Location Models” - textbook, excellent place to start • Three “state of the art” survey books • P. Mirchandani, R. Francis, 1990, “Discrete Location Theory” • Z. Drezner, 1995, “A Survey Of Applications And Methods” • Z. Drezner, H. Hamachar, 2004, “Location Theory: A survey of Applications and Methods” • New volume in the works… • Also of Interest • S. Nickel, J. Puerto, 2005, “Location Theory: A Unified Approach” – good reference for planar models • V. Marianov, H.A. Eiselt, 2011, “Foundations of Location Analysis” • Vast literature in various OR, OM, IE, Geography, Regional Science journals

  9. Overview • Introduction • Facility Location Problems – a quick review • Partial Capture Models: a Unifying framework • Modeling design aspects • Single-Facility Design Problem • Non-linear knapsack approach • Sensitivity analysis • Multi-facility Design and Location Problem • Tangent Line Approximation (TLA) approach to non-linear knapsack-type problems • Iterated TLA method • Conclusions and Future Research

  10. Goal • Want to model customer choice endogenously • Model should be realistic • Partial capture: good record of applications • Want to capture two key effects • Cannibalization • “Category expansion” • Need to model elastic demand • Need to incorporate facility “attraction” • Need a way to capture design elements • Start with a static model • Complex enough!

  11. Static Location and Design ModelsIncomplete literature review Full-Capture Models (deterministic customer choice) • MAXCAP: Revelle (1986), (…) • Location and Design Models • Plastria (1997), Plastria and Carrizosa (2003) – deterministic customer choice setting on a plane • Eiselt and Laporte (1989) – one facility, constant demand Partial-capture models (“discrete choice models”, “logit models”, “market share games”, etc.) • Spatial Interaction Models • Huff (1962, 1964), Nakanishi and Cooper (1974), (…), Berman and Krass (1998) • Spatial Interaction Models with Elastic Demand • Berman and Krass (2002), Aboolian, Berman, Krass (2006) - TLA • Competitive Facility Location and Design Problem (CFDLP) • Scenario design: Aboolian, Berman, Krass (2007) • Optimal design: Aboolian, Berman, Krass (??)

  12. Facility Location and Design ProblemModel Structure Objective (profit): (Total Captured Demand) - (Total Cost) Customer Demand: % captured by Facility j (customer choice) Di MSij Demand Customer Utility: Ui Utility of facility j for customer i Overall utility uij Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  13. Model Components: Facility Decisions Location Decisions • Discrete set of potential locations M • Competitive facilities may be present: set C • Must choose subset SM-C, |S|≤m • Binary decision variables xj=1 if location j chosen • Customers located at discrete set of points N • d(i,j) = distance from i to j • Fixed location cost fj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  14. Model Components: Facility Decisions Design Decisions • Attractiveness of facility at location j is given by • Assume design characteristics indexed by k=1,…,K • Typical characteristics: size, signage, #parking spaces, etc • j – attractiveness of “basic” (unimproved) facility at j • yjk – value of “improvement” of the facility with respect to k-th design characteristics • yk{0,1} for qualitative design characteristics • Log-linear form agrees with many marketing models; note concavity Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  15. Model Components: Facility Decisions Design Decisions • Attractiveness of facility at location j is given by • Cost: linear in decision variables Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  16. Model Components: UtilityUtility of facility j for customer i: uij • uij(Aj, d(i,j)) • Non-decreasing in attractiveness Aj • Decreasing in distance d(i,j) Customer Utility: Utility of facility j for customer i uij Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  17. Model Components: UtilityUtility of a given facility: Functional Form • Log-linear • Used in spatial interaction models • Exponential form is equivalent • Other functional forms can also be used Customer Utility: Utility of facility j for customer i uij Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  18. Model Components: UtilityOverall Utility Ui • uij(Aj, d(i,j)) • Uiis non-decreasing in uij for all i,j • Used Sum form: Customer Utility: Ui Utility of facility j for customer i Overall utility uij Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  19. Facility Location and Design ProblemPercent of Realized Customer Demand: Gi - • Gi(Ui) – non-negative, non-decreasing, concave function of total utility; 0≤ Gi(Ui)≤ 1 • D(Ui)= wiGi Customer Demand: % captured by Facility j (customer choice) Di MSij Demand Customer Utility: Ui Utility of facility j for customer i uij Overall utility Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  20. Model Components:Customer Demand • Gi(Ui) – non-negative, non-decreasing, concave function of total utility • 0≤G(Ui)≤1 represents realized proportion of potential demand from node i • wi- the maximum potential demand at i • Can write • Examples • Exponential demand: • Inelastic demand:

  21. Facility Location and Design ProblemPercent of Realized Customer Demand: Gi - • Spatial Interaction Models: Customer Demand: % captured by Facility j (customer choice) Di MSij Demand Customer Utility: Ui Utility of facility j for customer i uij Overall utility Travel distance d(i,j) Attractiveness Aj Facility Decisions: m xj yjk Number of facilities Locations Design Characteristics

  22. Model ComponentsMarket Share • Spatial Interaction Models: • note that total utility includes competitive facilities • also known as (or equivalent to) “logit”, “discrete choice”, “market-share games”, etc. • Full-capture model: • Total value Vi of customer i if facilities located in set S:

  23. Competitive Facility Location and Design Problem (CDFLP) Maximize total captured demand Design definition (attractiveness) The budgetary constraint Cannot improve unopened facility Select facility set S and design variables y

  24. Unifying Framework • This model unifies • Full and partial capture models • Constant / Elastic demand models • Models with / without design characteristics • General model very hard to solve directly • Non-linear IP; non-linearities in constraints and objective • Solvable cases • Constant demand, constant design (1998, 2002) • Elastic demand, constant design (2006) • Elastic demand, scenario design (2007) • General case: today

  25. Example • Assume a line segment network • No competitive facilities: Ui(C)=0 • Basic attractiveness j = 1 for j=1,2 • Only one design characteristic • yj = 2 or 0 (large or small facility) •  = .9 (large facility is 2.8x more attractive) • Budget allows us to locate two “small” or one “large” facility • Elasticity and distance sensitivity are set at 1 • ==1 distance =1 1 2 w1=1 w2=1

  26. distance =1 1 2 w1=1 w2=1 Illustration:Expansion and Cannibalization Demand Captured • Note that addition of second facility at 2 improved “company” picture, but not necessarily facility 1’s outlook – cannibalization and expansion in action First consider 1 small facility at node 1 Market shares Now add a second small facility at node 2

  27. Market Expansion vs. Cannibalization • Theorem: • Consider facility jX and customer iN • Suppose Gi(U) is concave • Let U.j= X-{j}uik+Ui(C) – utility derived by i from all other facilities • Let Dij(U.j) be the demand from i captured by j viewed as a function of U.j • Then Dij( ) is strictly decreasing in U.j • Implications: • Any improvements by other facilities (better design and/or new facilities by self or competitor) will reduce demand captured at facility j • Cannibalization effect always stronger than market expansion • Consequence of concave demand

  28. distance =1 distance =1 1 1 2 2 w1=1 w1=1 w2=1 w2=1 Corner stores vs. Supermarket Demand Captured • Here, one large facility performs slightly better Option 1: two small facilities Market shares Option 2: one large facility Market shares

  29. One “large” or two “small” facilities?Parametric Analysis – symmetric case Demand Elasticity Distance Sensitivity Design Sensitivity Conclusion: depending on sensitivity parameters, get either “corner store” or “supermarket” solutions

  30. One “large” or two “small” facilities?Competitive case (symmetric) • Assume locations are symmetric, but there are competitive facilities • U1(C) =2, U2(C) =1 (customers at 1 are better served by competition) Note that optimal location for large facility switches between 1 and 2 Why? Shouldn’t 2 be always preferred? Conclusion: depending on sensitivity parameters, get “corner store”, “box store”, or “mall” solutions – very flexible model

  31. CFDLP – Conceptual Solution Approach • Step 1: Solve 1-facility model for specified budget B • Equivalent to finding design characteristics that maximize attractiveness A for the given B • Solvable in closed form (non-linear knapsack) • Single-facility model can be solved by enumerating all potential facility locations • Step 2: Parametric analysis • Analyze A(B) optimal objective as a function of B • Can prove concavity; have quick algorithm for computing A(B) • Step 3: Back to multi-facility case

  32. Step 1: Single-Facility Design Problem(Index j suppressed) • Non-linear concave knapsack problem • Bretthauer and Shetty (EJOR, 2002); Birtran and Hax (MS, 1981) • Optimal solution can be computed in O(K2) time • Three sets: L, U, K-L-U • Characteristics in L “pegged” to LB of 0, • Characteristics in U pegged to the UB • Closed-form solution for all others Optimal Solution:

  33. Step 2: Parametric Analysis to Derive A(B) • For fixed sets L,U, K-L-U, can obtain a closed-form expression of optimal attractiveness as a function of the budget A*(B) • Optimal attractiveness is concave and non-decreasing in B • However, as B changes, so do sets L(B) and U(B) • Can identify (through linear search) a finite set of budgetary breakpoints B1,…BD • For B[Bb, Bb+1], set L(B) and U(B) are invariant and A*(B) is concave, non-decreasing in B • As B crosses a breakpoint, the slope of A*(B)changes • Can prove A*j(B) is concave, continuous and non-decreasing

  34. Parametric Analysis - Example • Theorem: A*(B) function is always concave(the derivative is discontinuous at breakpoints) K=3 B[3.5, 7] B B1=3.5 L= {2,3} U= B1=4 L={3}, U= B1=5 L=, U={1}

  35. CFDLP – Conceptual Solution Approach (cont) • Step 1: Solve 1-facility model for specified budget B • Step 2: Parametric analysis, derive A(B) • Step 3: Back to multi-facility case • All design variables yjk replaced with a single budget variable Bj Still difficult, but much more tractable non-linear IP Has knapsack-type structure Can prove that objective is a concave "superposition" of univariate concave functions

  36. CFDLP – Conceptual Solution Approach (cont) • Step 1: Solve 1-facility model for specified budget B • Step 2: Parametric analysis, derive A(B) • Step 3: Back to multi-facility case: replace design variables with Bj • Step 4: “Iterated TLA” • Utility Uiis separable with respect to A(Bj), concave • Objective function V(Ui) is also concave, composition of a concave function and a sum of univariate concave functions • Can apply a generalization of Tangent Line Approximation (TLA) method developed in Aboolian, Berman, Krass (2006) • Allows us to approximate the non-linear problem with a linear MIP • Approximation accuracy controllable by the user

  37. Tangent Line Approximation (TLA) Approach for a Class of Non-Linear Programs • Theorem (TLA): for any i and ε>0 can construct (in polynomial time) a piece-wise linear function Gεi(u) such that Gi(u) ≤ Gεi(u) and ( Gεi(u) – Gi(u))/Gi(u) ≤ 1-ε • i.e., Gεi(u) is an over-approximator within specified error bound • Moreover, Gεi(u) has the minimal number of linear segments among all piece-wise linear approximators of this precision level • Corollary 1: TLA converts NLP above into an LP whose solution is at most ε away from that of the original model (if original model was non-linear IP, get a linear IP) • For our problem, i(x) is concave in the decision variable, need a second application of TLA: “iterated TLA” • Also results in a single linear IP • Gi( ) is a concave, non-decreasing function, • i(x) – linear functional

  38. Tangent Line Approximation – Main Idea piece-wise linear approximator max relative error

  39. General CDFLP - Algorithm • Step 1: For each potential location derive breakpoints of A(B) • O(|K|2|M|) time • Step 2: Apply TLA approach to get piece-wise linear approximation • Polynomial approximation scheme • Step 3: Solve linear MIP • Size depends on solution accuracy set by the user

  40. Conclusions and Future Research • Very general and flexible framework • Single-facility location and design problem easy • Multi-facility problem tractable • Concave demand problem solvable through “iterated TLA” • Dimensionality grows over the regular TLA, but not too rapidly • Open Problems • Does the same methodology apply to “all or nothing” models • Dynamic competition

More Related