The Interaction of Location and Inventory in Designing Distribution Systems Stephen J. Erlebacher and Russell D. Meller Presented By: Hakan Gultekin Aim: • # of DCs • Their location • Which customers they serve Many DCs: • Reduces the cost of transporting product to retailers • Provide better service Few DCs: • Reduces the cost of holding inventory via pooling effects • Reduces the fixed costs associated with operating DCs The Location-Inventory Problem: Assumptions • Unit-square grid structure with C columns and R rows C= 5 1 2 3 4 5 1 2 3 4 R= 4 The Location-Inventory Problem: Assumptions • Uniform customer demand across any grid 19 ... 10 11 9 7 The Location-Inventory Problem: Assumptions • Rectilinear distances between plants and DCs and between DCs and continuously represented customer locations . (x,y) . (a,b) d x a y b The Location-Inventory Problem: Assumptions • Continuous review inventory system at DCs • Plant locations and capacities are known in advance and fixed The Location-Inventory Problem: The Model Minimize Total Cost where, Total Cost = (Operating Cost) + (DC Inventory Cost) + (Transportation Cost) The Location-Inventory Problem: The Model Operating Cost: F = Annual cost of operating a DC 1 if DCi is opened, zi 0 otherwise = Upper bound on the number of DCs F zi i 1 The Location-Inventory Problem: The Model Total Inventory Costs: A= Order cost z= Safety stock parameter h= Holding cost s Std. dev. of demand during lead time For DC i, order cost + holding cost: ADi Q 2 ADi h( zs L ) , Q Q 2 h M Di d j wij j 1 dj= Avg. demand for customer j 1 if DCi serves customer grid j, wij 0 otherwise The Location-Inventory Problem: The Model Total Inventory Costs: For 1 DC: zs 2 Ah D h M d w j 1 j ij For all DCs: zs zi 2 Ah h D i 1 M d w j 1 j ij The Location-Inventory Problem: The Model Transportation cost from plants to DCs: = Unit plant to DC transportation cost upi = Demand shipped from plant p to DC i qpi = Distance from plant p to DC i xi a p yi b p From plant p to DC i: zi u pi q pi From all plants to all DC s: P z u p 1 i 1 i pi q pi The Location-Inventory Problem: The Model Transportation cost from DCs to customers: tij = Avg distance from DCi to customer grid j 1 1 ( x (c j 1))(c j x) 2 2 . (c -1,y) j . . (x,y) (c ,y) j 1 1 ( x (c j 1)) ( x (c j 1)) (c j x) (c j x) 2 2 The Location-Inventory Problem: The Model Transportation cost from DCs to customers: tij = Avg distance from DCi to customer grid j 1 (x c j ) 2 1 2 . (c ,y) j 1 ( x c j(() c j 1) x) 2 . . (c ,y) (x,y) 1 1 x (c j 1) c j x 2 2 j 1 ((c j 1) x) 2 The Location-Inventory Problem: The Model Transportation cost from DCs to customers: s = Unit DC to customer transportation cost From DC i customer j: szi (t ij t ij )d j wij x y For all DCs and all customers: M x y sz ( t t i ij ij )d j wij i 1 j 1 The Location-Inventory Problem: The Model Constraints :: Each customer must be assigned to an open DC: M w ij j 1 Mzi , i One customer can be assigned to one DC: w i 1 ij 1, j The Location-Inventory Problem: The Model Constraints : Each DC must be fully supplied: P M u d w , p 1 pi j 1 j ij i Capacity constraint for plants : P u p 1 pi vp , p The Location-Inventory Problem: The Model zs F zi + zi 2 Ah D i 1 i 1 Min P + zi u pi q pi p 1 i 1 s.t. M w ij j 1 w i 1 ij + h 1, j i d w j j 1 ij M x y sz ( t t i ij ij )d j wij i 1 j 1 P Mzi , M M u d w , p 1 pi P u p 1 pi j 1 vp , j ij p zi {0,1}, i; wij {0,1}, i, j; u pi 0, p, i i The Location-Inventory Problem: Solution Method • Find the number of DCs, N • Find the location of these DCs and allocation of the customers to these DCs The Location-Inventory Problem: Solution Method Finding N: Stylized model • Customer demand is entirely homogeneos • Any amount of demand can be assigned to any DC • Each DC serves an “optimally shaped region” (for rectilinear, diamond shaped region) • Ignores different customer demands • Discrete nature of the customer grid structure • Impossible to have each DC serve an “optimally shaped region” The Location-Inventory Problem: Solution Method Finding N: Stylized model Lemma 1: Given a number of DCs, N, any DCs that serve positive demand must serve the same size demand. Di= D/N The Location-Inventory Problem: Solution Method Finding N: Stylized model Let I 2 ADh zs h be the inventory parameter, 2sD M T 3 2 V uD be the transportation parameter and be the inbound logistics costs. The Location-Inventory Problem: Solution Method Finding N: Stylized model Lemma 2: Optimal N for stylized model can be found by (i) 4F 2 N 3 I 2 N 2 2(T V )I N (T V )2 0 (ii) If N P, then N* N; Stop. (iii) 4[ F (V / P3/ 2 )]2 N 3 I 2 N 2 2T I N T 2 0 (iv) If N P, then N* P; else N * N. Optimal number of DCs The Location-Inventory Problem: Solution Method Actual Stylized Inventory Parameter Optimal number of DCs The Location-Inventory Problem: Solution Method Actual Stylized Transportation Parameter The Location-Inventory Problem: Solution Method Optimal number of DCs Actual Stylized Fixed Cost The Location-Inventory Problem: Solution Method Location Problems: • N-facility location problem NP-hard • N independent single facility location • Rectilinear mini-sum location problem The Location-Inventory Problem: Solution Method Allocation Heuristics v1: The Location-Inventory Problem: Solution Method Allocation Heuristics v2: Lower bound Relaxations: Separate inventory and transportation decisions Relax the actual customer locations Lemma 3: Lower bound on inventory costs are obtained by assigning the N-1 lowest demand customer grids to the first N-1 DCs and the remaining M-N+1 customer girds to DCN Sort customers from highest demand to lowest demand and assign them one at a time to a DC. Highest demand customers have more influence on the location of the DC which they are assigned Computational Results & Managerial Insight Two datasets considered: • Set I consists of 12 customers (on a 3X4 grid) • Set II consists of 16 customers (on a 4X4 grid) • 4 different ABC customer curves: (80/20), (70/30), (60/40), (50/50) • v2 performed better than v1 in both sets Computational Results & Managerial Insight • Lower bound was between 4 and 36 % lower than the optimal solution • Neither of the heuristics are guaranteed to terminate at a local optimum. • Pairwise-exchange improvement procedure is added (v2+). • For dataset I, v2 found the optimum in 10/75 while v2+ found in 62/75 • For 600 customers v2 solved in 2 minutes, v1 solved in 30 hours and v2+ solved in 117 hours. Computational Results & Managerial Insight • As the skewness of ABC curve increases, N either stays the same or decreases, since larger demand is concentrated in fewer and fewer customers. • The customer layout also affects the optimal number of DCs. • Geary Ratio, is an autocorrelation factor that quantifies spatial correlations • Tends to decrease when similarly-sized customer demands are adjacent Computational Results & Managerial Insight . . . . . 0.25 0.27 0.28 . . 0.23 . . 0.26 . . 0.39 . . . . . . 0.60 0.33 . 41.5 1.76 . 0.29 0.74 . 1.08 0.41 0.23 411.9 0.31 0.63 . . . . . . . . . . . . . . . 0.25 0.27 0.28 . . 0.23 . . 0.26 . . 0.39 . . . . . . 0.60 0.33 . 0.23 1.76 . 0.29 0.74 . 1.08 0.410 41.5 411.9 0.31 0.63 . . . . . . . . . . • Higher Geary Ratio • Smaller Geary Ratio • Higher number of DCs • Smaller number of DCs Future Research • Capacity limitations at the DCs • Different type of inventory policies • Multi-product environment The Interaction of Location and Inventory in Designing Distribution Systems Stephen J. Erlebacher and Russell D. Meller Presented By: Hakan Gultekin