Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location...
Transcript of Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location...
![Page 1: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/1.jpg)
Integrating Facility Location andNetwork Design
Amitabh Sinha
(Joint work with R. Ravi)
GSIA, Carnegie Mellon University
Integrating Facility Location and Network Design – p.1
![Page 2: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/2.jpg)
Motivation: UFL with cable capacities
• Facility Location
with cable ca-pacities
• Input: Set of clients & facilities(with opening costs) in a metricspace.
• Objective: Open some facilitiesto serve clients. Client servicecost: distance to nearest openfacility. Minimize total cost.
• New twist: Clients connect tofacilities via (capacitated) ca-bles. Service cost becomesmore complicated.
INPUT
Clients Facilities
Integrating Facility Location and Network Design – p.2
![Page 3: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/3.jpg)
Motivation: UFL with cable capacities
• Facility Location
with cable ca-pacities
• Input: Set of clients & facilities(with opening costs) in a metricspace.
• Objective: Open some facilitiesto serve clients. Client servicecost: distance to nearest openfacility. Minimize total cost.
• New twist: Clients connect tofacilities via (capacitated) ca-bles. Service cost becomesmore complicated.
INPUT
Clients Facilities
Integrating Facility Location and Network Design – p.2
![Page 4: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/4.jpg)
Motivation: UFL with cable capacities
• Facility Location
with cable ca-pacities
• Input: Set of clients & facilities(with opening costs) in a metricspace.
• Objective: Open some facilitiesto serve clients. Client servicecost: distance to nearest openfacility. Minimize total cost.
• New twist: Clients connect tofacilities via (capacitated) ca-bles. Service cost becomesmore complicated.
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.2
![Page 5: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/5.jpg)
Motivation: UFL with cable capacities
• Facility Location with cable ca-pacities
• Input: Set of clients & facilities(with opening costs) in a metricspace.
• Objective: Open some facilitiesto serve clients. Client servicecost: distance to nearest openfacility. Minimize total cost.
• New twist: Clients connect tofacilities via (capacitated) ca-bles. Service cost becomesmore complicated.
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.2
![Page 6: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/6.jpg)
Motivation: UFL with cable capacities
• Facility Location with cable ca-pacities
• Input: Set of clients & facilities(with opening costs) in a metricspace.
• Objective: Open some facilitiesto serve clients. Client servicecost: distance to nearest openfacility. Minimize total cost.
• New twist: Clients connect tofacilities via (capacitated) ca-bles. Service cost becomesmore complicated.
Cable capacity = 3
CCFL: feasible solution
Integrating Facility Location and Network Design – p.2
![Page 7: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/7.jpg)
Outline
• Define CCFL: Capacitated cable facility location.
• Lower bounds for CCFL.
• Approximation algorithm for CCFL.
• Define KCFL: k-cable facility location.
• Thoughts on approximating KCFL.
Integrating Facility Location and Network Design – p.3
![Page 8: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/8.jpg)
Problem definition
• Capacitated Cable Facility Lo-cation (CCFL):
• Graph (metric), Edge weightsce, Clients D ⊆ V , Facilities Fwith costs φj , and Cable capac-ity u.
• Goal: Open some facilities, andinstall cables on edges, to sup-port 1 unit of flow from eachclient to some open facility.
• Objective: Minimize total cost(facilities + cables).
INPUT
Clients Facilities
Integrating Facility Location and Network Design – p.4
![Page 9: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/9.jpg)
Problem definition
• Capacitated Cable Facility Lo-cation (CCFL):
• Graph (metric), Edge weightsce, Clients D ⊆ V , Facilities Fwith costs φj , and Cable capac-ity u.
• Goal: Open some facilities, andinstall cables on edges, to sup-port 1 unit of flow from eachclient to some open facility.
• Objective: Minimize total cost(facilities + cables).
INPUT
Clients Facilities
Integrating Facility Location and Network Design – p.4
![Page 10: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/10.jpg)
Problem definition
• Capacitated Cable Facility Lo-cation (CCFL):
• Graph (metric), Edge weightsce, Clients D ⊆ V , Facilities Fwith costs φj , and Cable capac-ity u.
• Goal: Open some facilities, andinstall cables on edges, to sup-port 1 unit of flow from eachclient to some open facility.
• Objective: Minimize total cost(facilities + cables).
INPUT
Clients Facilities
Integrating Facility Location and Network Design – p.4
![Page 11: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/11.jpg)
Problem definition
• Capacitated Cable Facility Lo-cation (CCFL):
• Graph (metric), Edge weightsce, Clients D ⊆ V , Facilities Fwith costs φj , and Cable capac-ity u.
• Goal: Open some facilities, andinstall cables on edges, to sup-port 1 unit of flow from eachclient to some open facility.
• Objective: Minimize total cost(facilities + cables).
Cable capacity = 3
CCFL: feasible solution
Integrating Facility Location and Network Design – p.4
![Page 12: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/12.jpg)
Special cases and past work
• u = 1: UFL; ρUFL = 1.52 [MYZ02]. Others: [STA 97, JV 99,AGKMMP 01, JMS 02].
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.5
![Page 13: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/13.jpg)
Special cases and past work
• u = 1: UFL; ρUFL = 1.52 [MYZ02]. Others: [STA 97, JV 99,AGKMMP 01, JMS 02].
• u = ∞: Steiner tree; ρST = 1.55[RZ 99]. Others: [TM 80, AKR95, Zel 95, HP 99].
Steiner tree
Clients Open Facilities
Integrating Facility Location and Network Design – p.5
![Page 14: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/14.jpg)
Special cases and past work
• u = 1: UFL; ρUFL = 1.52 [MYZ02]. Others: [STA 97, JV 99,AGKMMP 01, JMS 02].
• u = ∞: Steiner tree; ρST = 1.55[RZ 99]. Others: [TM 80, AKR95, Zel 95, HP 99].
• |F| = 1: Single sink single ca-ble edge installation; ρSS = 3[HRS 00]. Others: [AA 97, AZ98, GKKRSS 01, GMM 01, Tal02].
Clients Sink
Single sink edge installation
Integrating Facility Location and Network Design – p.5
![Page 15: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/15.jpg)
Special cases and past work
• u = 1: UFL; ρUFL = 1.52 [MYZ02]. Others: [STA 97, JV 99,AGKMMP 01, JMS 02].
• u = ∞: Steiner tree; ρST = 1.55[RZ 99]. Others: [TM 80, AKR95, Zel 95, HP 99].
• |F| = 1: Single sink single ca-ble edge installation; ρSS = 3[HRS 00]. Others: [AA 97, AZ98, GKKRSS 01, GMM 01, Tal02].
• CCFL: O(log n) [MMP 00], alsofor KCFL. This paper: 3.07
Cable capacity = 3
CCFL: feasible solution
Integrating Facility Location and Network Design – p.5
![Page 16: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/16.jpg)
Lower bound: Routing
• New UFL instance: Scale edgecosts to c′
e = ce/u.
• OPT (UFL) ≤ OPT (CCFL).
• Reason: In CCFL, each clientincurs service cost at least 1/uof the cost of its path to its facil-ity.
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.6
![Page 17: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/17.jpg)
Lower bound: Connectivity
• New Steiner tree instance: Addroot r, connect to each facilitywith edge cost φj . Terminals:D ∪ {r}.
• OPT (ST ) ≤ OPT (CCFL).
• Reason: In CCFL, each clientmust have a connection to somefacility.
Steiner tree
Clients Open Facilities
Integrating Facility Location and Network Design – p.7
![Page 18: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/18.jpg)
Algorithm motivation
• Routing LB: Good for high de-mand, bad for low demand.
• Connectivity LB: Bad for highdemand, good for low demand.
• How to combine them?
• Use ideas from single sinkedge installation algorithm!
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.8
![Page 19: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/19.jpg)
Algorithm motivation
• Routing LB: Good for high de-mand, bad for low demand.
• Connectivity LB: Bad for highdemand, good for low demand.
• How to combine them?
• Use ideas from single sinkedge installation algorithm!
Steiner tree
Clients Open Facilities
Integrating Facility Location and Network Design – p.8
![Page 20: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/20.jpg)
Algorithm motivation
• Routing LB: Good for high de-mand, bad for low demand.
• Connectivity LB: Bad for highdemand, good for low demand.
• How to combine them?
• Use ideas from single sinkedge installation algorithm!
Clients Sink
Single sink edge installation
Integrating Facility Location and Network Design – p.8
![Page 21: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/21.jpg)
Algorithm description ... 1
1. Solve scaled UFL (c′
e = ce/u).
2. Solve Steiner tree instance.
3. Open facilities of both stages.Install cables of Steiner treestage.
This is infeasible!
4. Convert to feasible solution byaggregating demand and in-stalling new cables.
(Details coming up.)
UFL Solution
Clients Open Facilities
Integrating Facility Location and Network Design – p.9
![Page 22: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/22.jpg)
Algorithm description ... 1
1. Solve scaled UFL (c′
e = ce/u).
2. Solve Steiner tree instance.
3. Open facilities of both stages.Install cables of Steiner treestage.
This is infeasible!
4. Convert to feasible solution byaggregating demand and in-stalling new cables.
(Details coming up.)
Steiner tree
Clients Open Facilities
Integrating Facility Location and Network Design – p.9
![Page 23: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/23.jpg)
Algorithm description ... 1
1. Solve scaled UFL (c′
e = ce/u).
2. Solve Steiner tree instance.
3. Open facilities of both stages.Install cables of Steiner treestage.
This is infeasible!
4. Convert to feasible solution byaggregating demand and in-stalling new cables.
(Details coming up.)
Algorithm: Step 3.
Clients Open Facilities
Integrating Facility Location and Network Design – p.9
![Page 24: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/24.jpg)
Algorithm description ... 1
1. Solve scaled UFL (c′
e = ce/u).
2. Solve Steiner tree instance.
3. Open facilities of both stages.Install cables of Steiner treestage.
This is infeasible!
4. Convert to feasible solution byaggregating demand and in-stalling new cables.
(Details coming up.)Algorithm: Step 3.
Clients Open Facilities
Integrating Facility Location and Network Design – p.9
![Page 25: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/25.jpg)
Algorithm description ... 2
4. Installing new cables to makesolution feasible:
For each tree in forest:(a) Identify “lowest” node with
demand ≥ u.Form “clump” of u nodes insuch a subtree.
(b) In this “clump”, install anew cable to connect near-est client-facility pair.
(c) Reroute flow appropriately.
Algorithm: Step 3.
Clients Open Facilities
Integrating Facility Location and Network Design – p.10
![Page 26: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/26.jpg)
Algorithm description ... 2
4. Installing new cables to makesolution feasible:For each tree in forest:
(a) Identify “lowest” node withdemand ≥ u.Form “clump” of u nodes insuch a subtree.
(b) In this “clump”, install anew cable to connect near-est client-facility pair.
(c) Reroute flow appropriately.
Algorithm: Step 4(a).
Clients Open Facilities
1
2 1
4
5
Integrating Facility Location and Network Design – p.10
![Page 27: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/27.jpg)
Algorithm description ... 2
4. Installing new cables to makesolution feasible:For each tree in forest:
(a) Identify “lowest” node withdemand ≥ u.Form “clump” of u nodes insuch a subtree.
(b) In this “clump”, install anew cable to connect near-est client-facility pair.
(c) Reroute flow appropriately.
Clients Open Facilities
Algorithm: Step 4(b).
1
2
11
2
3
Integrating Facility Location and Network Design – p.10
![Page 28: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/28.jpg)
Algorithm description ... 2
4. Installing new cables to makesolution feasible:For each tree in forest:
(a) Identify “lowest” node withdemand ≥ u.Form “clump” of u nodes insuch a subtree.
(b) In this “clump”, install anew cable to connect near-est client-facility pair.
(c) Reroute flow appropriately.
Clients Open Facilities
Algorithm: Step 4(b).
1
2
11
2
3
Integrating Facility Location and Network Design – p.10
![Page 29: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/29.jpg)
Algorithm description ... 2
4. Installing new cables to makesolution feasible:For each tree in forest:
(a) Identify “lowest” node withdemand ≥ u.Form “clump” of u nodes insuch a subtree.
(b) In this “clump”, install anew cable to connect near-est client-facility pair.
(c) Reroute flow appropriately.
Clients Open Facilities
Algorithm: Final solution.
Integrating Facility Location and Network Design – p.10
![Page 30: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/30.jpg)
Performance analysis
• Theorem [HRS 00]: Aggregation-and-reroutingproduces feasible solution.
• Facility cost: Paid by the two lower bounds.
• Cables on Steiner tree: Paid by Steiner tree lowerbound.
• New cables from “clumps”: Paid by routing (service)cost component of UFL solution, since each client inUFL solution incurs c/u service cost and each clumphas u clients.
• Theorem: The algorithm is a ρST + ρUFL (≈ 3.07)approximation for CCFL.
Integrating Facility Location and Network Design – p.11
![Page 31: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/31.jpg)
Performance analysis
• Theorem [HRS 00]: Aggregation-and-reroutingproduces feasible solution.
• Facility cost: Paid by the two lower bounds.
• Cables on Steiner tree: Paid by Steiner tree lowerbound.
• New cables from “clumps”: Paid by routing (service)cost component of UFL solution, since each client inUFL solution incurs c/u service cost and each clumphas u clients.
• Theorem: The algorithm is a ρST + ρUFL (≈ 3.07)approximation for CCFL.
Integrating Facility Location and Network Design – p.11
![Page 32: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/32.jpg)
Performance analysis
• Theorem [HRS 00]: Aggregation-and-reroutingproduces feasible solution.
• Facility cost: Paid by the two lower bounds.
• Cables on Steiner tree: Paid by Steiner tree lowerbound.
• New cables from “clumps”: Paid by routing (service)cost component of UFL solution, since each client inUFL solution incurs c/u service cost and each clumphas u clients.
• Theorem: The algorithm is a ρST + ρUFL (≈ 3.07)approximation for CCFL.
Integrating Facility Location and Network Design – p.11
![Page 33: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/33.jpg)
Performance analysis
• Theorem [HRS 00]: Aggregation-and-reroutingproduces feasible solution.
• Facility cost: Paid by the two lower bounds.
• Cables on Steiner tree: Paid by Steiner tree lowerbound.
• New cables from “clumps”: Paid by routing (service)cost component of UFL solution, since each client inUFL solution incurs c/u service cost and each clumphas u clients.
• Theorem: The algorithm is a ρST + ρUFL (≈ 3.07)approximation for CCFL.
Integrating Facility Location and Network Design – p.11
![Page 34: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/34.jpg)
Performance analysis
• Theorem [HRS 00]: Aggregation-and-reroutingproduces feasible solution.
• Facility cost: Paid by the two lower bounds.
• Cables on Steiner tree: Paid by Steiner tree lowerbound.
• New cables from “clumps”: Paid by routing (service)cost component of UFL solution, since each client inUFL solution incurs c/u service cost and each clumphas u clients.
• Theorem: The algorithm is a ρST + ρUFL (≈ 3.07)approximation for CCFL.
Integrating Facility Location and Network Design – p.11
![Page 35: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/35.jpg)
Thoughts
• Natural IP formulation has good relaxation. Ouralgorithm yields a gapST + gapUFL (≈ 5) LP-roundingapproximation algorithm for CCFL.
• Generalizes to non-uniform demands at clients. Ifdemand is splittable, performance ratio remains same(≈ 3.07).For unsplittable demand, the aggregate-and-reroutestep needs a little more work. Performance ratio is nowρST + 2ρUFL (≈ 4.59).
• No tight example known.Lower bound on approximation ratio is 1.46, comingfrom UFL.
Integrating Facility Location and Network Design – p.12
![Page 36: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/36.jpg)
Thoughts
• Natural IP formulation has good relaxation. Ouralgorithm yields a gapST + gapUFL (≈ 5) LP-roundingapproximation algorithm for CCFL.
• Generalizes to non-uniform demands at clients. Ifdemand is splittable, performance ratio remains same(≈ 3.07).For unsplittable demand, the aggregate-and-reroutestep needs a little more work. Performance ratio is nowρST + 2ρUFL (≈ 4.59).
• No tight example known.Lower bound on approximation ratio is 1.46, comingfrom UFL.
Integrating Facility Location and Network Design – p.12
![Page 37: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/37.jpg)
Thoughts
• Natural IP formulation has good relaxation. Ouralgorithm yields a gapST + gapUFL (≈ 5) LP-roundingapproximation algorithm for CCFL.
• Generalizes to non-uniform demands at clients. Ifdemand is splittable, performance ratio remains same(≈ 3.07).For unsplittable demand, the aggregate-and-reroutestep needs a little more work. Performance ratio is nowρST + 2ρUFL (≈ 4.59).
• No tight example known.Lower bound on approximation ratio is 1.46, comingfrom UFL.
Integrating Facility Location and Network Design – p.12
![Page 38: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/38.jpg)
KCFL: k-cable facility location
• Variant of CCFL: k cable typesto choose from.
Depending on flow, one partic-ular type of cable may be mosteconomical.
• Current status:
O(log n) due to [MMP 00],
O(k) due to [RS 02].
Cable capacity = 3
CCFL: feasible solution
Integrating Facility Location and Network Design – p.13
![Page 39: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/39.jpg)
KCFL: k-cable facility location
• Variant of CCFL: k cable typesto choose from.
Depending on flow, one partic-ular type of cable may be mosteconomical.
• Current status:
O(log n) due to [MMP 00],
O(k) due to [RS 02].
KCFL: feasible solution
Integrating Facility Location and Network Design – p.13
![Page 40: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/40.jpg)
k-cable single sink edge installation
• Single sink: |F| = 1.
• O(1) approximation due to[GMM 01].Combinatorial, randomized al-gorithm, using same structurallower bounds (routing and con-nectivity).
Clients Sink
k−cable edge installation
Integrating Facility Location and Network Design – p.14
![Page 41: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/41.jpg)
k-cable single sink edge installation
• Single sink: |F| = 1.
• O(1) approximation due to[GMM 01].Combinatorial, randomized al-gorithm, using same structurallower bounds (routing and con-nectivity).
• Improved O(1) approximationdue to [Tal 02].LP rounding, improves on O(k)of [GKKRSS 01].
Clients Sink
k−cable edge installation
Integrating Facility Location and Network Design – p.14
![Page 42: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/42.jpg)
Thoughts on approximating KCFL
• Open: O(1) for KCFL?
Integrating Facility Location and Network Design – p.15
![Page 43: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/43.jpg)
Thoughts on approximating KCFL
• Open: O(1) for KCFL?
• Extend [GMM 01]: yields O(k) [RS 02], with O(1) oncable cost. O(k) is only due to facility costs.Better algorithm / analysis may yield O(1).
Integrating Facility Location and Network Design – p.15
![Page 44: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/44.jpg)
Thoughts on approximating KCFL
• Open: O(1) for KCFL?
• Extend [GMM 01]: yields O(k) [RS 02], with O(1) oncable cost. O(k) is only due to facility costs.Better algorithm / analysis may yield O(1).
• Slight modification of LP of [Tal 02] yields formulation ofKCFL.Open: Rounding or gap for LP.
Integrating Facility Location and Network Design – p.15
![Page 45: Integrating Facility Location and Network Design · 2010-11-17 · Integrating Facility Location and Network Design Amitabh Sinha (Joint work with R. Ravi) GSIA, Carnegie Mellon University](https://reader034.fdocuments.us/reader034/viewer/2022042411/5f29e646a492bc65264cdcb3/html5/thumbnails/45.jpg)
Thoughts on approximating KCFL
• Open: O(1) for KCFL?
• Extend [GMM 01]: yields O(k) [RS 02], with O(1) oncable cost. O(k) is only due to facility costs.Better algorithm / analysis may yield O(1).
• Slight modification of LP of [Tal 02] yields formulation ofKCFL.Open: Rounding or gap for LP.
• Open: CCFL / KCFL with capacitated facilities.
Integrating Facility Location and Network Design – p.15