25 min - Yisong Yue

An Efficient
Simulation-based Approach to Ambulance
Fleet Allocation and Dynamic Redeployment
Key questions:
25 min
(20 min)
-- If I place my ambulances
elsewhere, will that improve performance?
18 min
minutes
saved)
-- How
do I do automate (12
this
process?
25 min
30 min
Data-driven simulation:
Optimization
Algor ithm
: Request Sampler as subroutine
• 2 SUse simulator
1: input: t
,t
2: R ← •
∅, t ←
t
We
use a greedy approach
3: while t < t
do
4:
Sample r ← P (r |t) //sampling request starting at time t
• Suitable
5:
t←t
//incrementing
current time for real-time redeployment
A M PL ER
st a r t
(17m)
en d
st a r t
en d
r
(15 min)
6:
R ← R ∪ {r}
7: end while
8: retur n: R
//adding sampled request to collection
Analysis
• Objective function hard to analyze directly
This independence between request arrival patterns and
• Generative model of requests EMS
25
min
behavior
will beprove
important when
developing our
opti•
We
data-dependent
guarantees
(25dispatch
min)
mization
approach
and
analysis.
When
simulating
• Simulate outcomes 15 min
(no
vehicle
assigned previously)
behavior (and evaluating EMS service levels), we can first
• a call
Show
greedy
near-optimal
pre-sample
log R (e.g.,
one week’sis
worth),
and then
45 min
• Measure metrics of(30
interest
minutes saved)
evaluate service quality of an allocation A by running S described in Algorithm 1. On the other hand, if reEmpirical
quest arrival
depends on theValidation
EMS allocation and dispatch
behavior (and thus violates Assumption 1), then one can no
longer pre-sample
futurestudy
calls for simulation.
• Case
on EMS system of Asian city
Figure 1: Examples of simulated dispatch of four requests
Assumption
1
also
implies
that R can be incrementally
and two bases. The top example has one ambulance allosampled •according
to a memory-less
stochastic process.
cated per base. The bottom example has two ambulances alSignificant
performance
gains
Thus, for any time interval (e.g., one week), we can increlocationed to A and one to A .
17m
IM
UL ATOR
Come see poster!
1
2
areperformed: r isremoved from theset of activerequests
(Line 17), and the assigned ambulance is added to the set
of available ambulances (Line 18).
3
mentally sample requests using Algorithm 2. Here, P (r |t)
denotes the distribution of the next arriving request starting
at time t, and t r denotes the arrival time of request r .
Yisong Yue (CMU) & Lavanya Marla (CMU) & Ramayya Krishnan (CMU)
4 Static Allocation Problem For mulation