deploying ambulances: some insights and computational tools · 2010-01-11 · parallel computing...
TRANSCRIPT
![Page 1: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/1.jpg)
Deploying Ambulances: SomeInsights and Computational Tools
Shane G. HendersonCornell University
Joint work with Mateo Restrepo,Huseyin Topaloglu
Thanks to NSF DMI 0400287
![Page 2: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/2.jpg)
Shane G. Henderson 2
Goal and Outline
• Goal: Develop insight and computational toolsfor ambulance-deployment questions
• Why do small towns have to run ambulancesat lower utilizations than cities?
• Should the number of ambulances inseparate regions be proportional to thenumber of calls?
• A New Approach to System-StatusManagement
![Page 3: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/3.jpg)
Shane G. Henderson 3
Why do small towns haveto run ambulances at lower
utilizations than cities?
![Page 4: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/4.jpg)
Shane G. Henderson 4
Number of Calls in a Shift
Typical Fluctuation = 3
![Page 5: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/5.jpg)
Shane G. Henderson 5
Number of Calls in a Shift
Typical Fluctuation = 8
![Page 6: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/6.jpg)
Shane G. Henderson 6
Number of Calls in a Shift
Typical Fluctuation = 16
![Page 7: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/7.jpg)
Shane G. Henderson 7
1664
816
32
FluctuationsAverage Numof Calls
Number of Calls in a Shift
N 2 √N
• Theory predicts#calls is roughlydistributed asPoisson(mean)
• For Poisson,std dev = √average
• As average gets big,relative fluctuationsget small
![Page 8: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/8.jpg)
Shane G. Henderson 8
Impact on Deployment
• Need to plan for N + 2 √N calls, whereN = average
• When N is small this is much biggerthan the mean
• When N is big, not much difference• Disclaimer: Assumes all ambulances on
shift can cooperate. Not always true inlarge areas
![Page 9: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/9.jpg)
Shane G. Henderson 9
Should the number ofambulances in separate
regions be proportional tothe number of calls?
![Page 10: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/10.jpg)
Shane G. Henderson 10
Allocating AmbulancesProvincial orregional planning
Traffic congestiondisconnects city
![Page 11: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/11.jpg)
Shane G. Henderson 11
• Two regions, no communication• Calls that arrive when all ambulances in the
region are busy are “lost”• What arrangment of c ambulances minimizes
lost calls?
![Page 12: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/12.jpg)
Shane G. Henderson 12
Two Locations, No Interaction
• Arrival rates λ1 and λ2
• Minimize λ1 L(λ1, c1) + λ2 L(λ2, c2)• s/t c1 + c2 = c
• Q: Should ci* be proportional to λi ?
• A: No…
![Page 13: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/13.jpg)
Shane G. Henderson 13
50 ambulances, 20 calls/hr
![Page 14: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/14.jpg)
Shane G. Henderson 14
10 ambulances, 4 Calls Per Hr
![Page 15: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/15.jpg)
Shane G. Henderson 15
So Where are We?
• Allocating in proportion to demand isnot always best
• Better to boost allocation to smallerlocation slightly
• Limitation: Assumes all ambulances ata location share the total load
![Page 16: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/16.jpg)
Shane G. Henderson 16
A New Approach toSystem-Status Management
![Page 17: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/17.jpg)
Shane G. Henderson 17
Relocation
AKA– Relocation– Redeployment– System status
management– Move up
![Page 18: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/18.jpg)
Shane G. Henderson 18
Existing SSM Methods
• Have lookup table: Ideal ambulancelocations given number available
• Dispatch available ambulances to keepthose locations full
• Issues:– Potentially many “cold” moves– Crew prefers base to street corners– How do we know it’s better?
![Page 19: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/19.jpg)
Shane G. Henderson 19
Ingredients for our Method
• Real-time information on ambulances– GPS coordinates– Status, e.g., at hospital for 12 minutes
• A simulation model of ambulance dynamics• A function, V say, that gives the value or
quality of a configuration• Managerial decisions on when and where to
consider redeployment– E.g., Upon call completion only, or other times too– E.g., Only to bases, or to other locations too
![Page 20: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/20.jpg)
Shane G. Henderson 20
![Page 21: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/21.jpg)
Shane G. Henderson 21
1
Assume send ambulance to 1
Do 10 times:
Simulate immediate future, and look up V for final ambulance positions, status
Compute average of V values
= 50.1 say
V=50.1
2
Assume send ambulance to 2
Do 10 times:
Simulate immediate future, and look up V for final ambulance positions, status
Compute average of V values
= 63.2
V=63.2
![Page 22: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/22.jpg)
Shane G. Henderson 22
Some Features
• Super-fast simulations in real time• Exploits upcoming information, e.g.,
Ambulance 2 should be free shortly• Can limit potential moves
– To ambulances completing a call, or– To ambulances already on the road, or– To small set of destinations
• Dispatcher can intervene - override, query• Can restrict relocations to “high benefit” ones
![Page 23: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/23.jpg)
Shane G. Henderson 23
Where Did V Come From?
• Selected in initial training phase - not inreal time
• Use “Approximate DynamicProgramming” methods in conjunctionwith simulation: “computer learning”
• Guess V - simulate it - improve it• Computationally intensive!
![Page 24: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/24.jpg)
Shane G. Henderson 24
Shortcomings
• Need a simulation model– Must be maintained– Be careful to use right model for right day
• Need to customize for each installation• Doesn’t work well with “large scale events”
like 9/11• Haven’t yet tried multiple levels of care
![Page 25: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/25.jpg)
Shane G. Henderson 25
“Convergence”(Edmonton – Thanks Armann)
![Page 26: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/26.jpg)
Shane G. Henderson 26
“Convergence” (Edm)
![Page 27: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/27.jpg)
Shane G. Henderson 27
Extra Redeployments? (Edm)
![Page 28: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/28.jpg)
Shane G. Henderson 28
Other Call Volumes?(Certain City - Thanks Alex, Andrew)
![Page 29: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/29.jpg)
Shane G. Henderson 29
Varying # Ambulances
![Page 30: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/30.jpg)
Shane G. Henderson 30
What is Next For Us?
• Multiple levels of care• How much better are street corners
than bases?• Interaction with large-scale events• Testing on high-resolution simulation
software
![Page 31: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/31.jpg)
Shane G. Henderson 31
References• Andersson and Värband. 2006. Decision support tools for ambulance dispatch and
relocation. Journal of Operational Research Society• Berman, O. 1981. Dynamic repositioning of indistinguishable service units on transportation
networks. Transportation Science, 15:115-136.• Bertsekas and Tsitsiklis 1996. Neuro-Dynamic Programming. Athena Scientific.• Brotcorne, Laporte and Semet. 2003. Ambulance location and relocation models. European
Journal of Operational Research, 147: 451-463.• Goldberg. 2004. Operations research models for the deployment of emergency services
vehicles. EMS Management Journal, 1:20-39.• Henderson and Mason. 2004. Ambulance service planning: simulation and data
visualization. In M. L. Brandeau, F. Sainfort, and W. P. Pierskalla, eds, OperationsResearch and Health Care: A Handbook of Methods and Applications, 77-102. KluwerAcademic, Boston.
• Ingolfsson and coauthors at Edmonton – many papers• Kolesar and Walker. 1974. An algorithm for the dynamic relocation of fire companies.
Operations Research, 22:249-274.• Laporte, Gendreau and Semet. 2001. A dynamic model and parallel tabu search heuristic
for real-time ambulance relocation. Parallel Computing 27:1641-1653.• Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering location model for
dynamic redeployment of ambulances. Manuscript.• Restrepo, Henderson and Topaloglu. 2008. Erlang loss models for the static deployment of
ambulances. Health Care Management Science. To appear.• Restrepo, Henderson and Topaloglu. 2008. Approximate dynamic programming for
ambulance redeployment. Manuscript.
![Page 32: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/32.jpg)
Shane G. Henderson 32
Wrap-up• Rural areas have higher ratio of
variability to average load• Should increase resources to lower-
loaded areas beyond proportional• Dynamic relocation can substantially
improve day-to-day performance• Approximate DP: some art required
![Page 33: Deploying Ambulances: Some Insights and Computational Tools · 2010-01-11 · Parallel Computing 27:1641-1653. • Rajagopalan, Saydam and Xiao. 2006. A multiperiod expected covering](https://reader034.vdocument.in/reader034/viewer/2022042222/5ec8dc8e14ea260bfd794e55/html5/thumbnails/33.jpg)
Shane G. Henderson 33
Some Basis Functions
• Expected # missed calls over remaininghorizon
• # waiting calls that will be missed• Rate of calls arriving to uncovered areas• Rate of calls arriving to covered areas, that
will likely (Erlang loss) be missed• If ambulances are not diverted from present
paths, future locations are also important.Future versions of last 2 basis functions