2. Why Reliability is important in Shortest Path Problems
Reliable Shortest Path Formulation
Outer Approximation Solution Algorithm
Computational Study
Conclusions and Other Applications
3. s t
35
45
STD DEV
MEAN TRAVEL COST
, 5
, 20
ORIGIN DESTINATION
Reliable and risk averse routing suggest taking
the path with least mean+std cost
TC=55
TC=50
5. Directed Graph G=(V,E)
Origin Node r, Destination Node s
aij is the link between node i and node j
cov(aij, alk) is the covariance matrix
Given
Assumptions
Link Travel Cost Distribution is Available
One path with finite variance
8. 𝝈𝒊𝒋
𝟐
𝒙𝒊𝒋
𝒓𝒔 𝟐
+ 𝒄𝒐𝒗 𝒂𝒊𝒋, 𝒂𝒍𝒌 𝒙𝒊𝒋
𝒓𝒔
𝒙𝒍𝒌
𝒓𝒔
𝒂 𝒍𝒌 ∈ 𝑬
𝒍,𝒌 ≠ 𝒊,𝒋
𝒂 𝒊𝒋 ∈ 𝑬𝒂 𝒊𝒋 ∈ 𝑬
≤ 𝒕 𝒓𝒔
Conic Formulation inefficient for Large Networks
Consider a network with 1000 links
1 million quadratic integer terms
9. CQP computationally tractable due to their special Structure
Can be solved by polynomial time interior point algorithms
Solvers like CPLEX and MOSEK offering this capability
X2 + Y2 <= Z2
.
Minimize
Subject to
ii
T
iiii
dxcbxa
i
T
i
xf
10. Outer Approximation (OA)-Linearization and Decomposition
Method
Delivers optimal solution for convex MINLP (Bonami et al.
2008)
ZyRx
yxg
yxf
yx
,
0),(toSubject
Minimize ),(
,
f and g are convex function
Decompose integer and nonlinear part
Solve alternating sequence of Master Problem
(MILP) & Subproblem (NLP)
17. Input: Convergence tolerance ε, Maximum Number of
Iteration H, Upperbound(UB)=+inf, Lowerbound (LB) =-
inf, h=1
Initialization: Solve One-all Shortest path problem
1: If (UB-LB) <=ε or then go to step 6
2: Solve the SP (Closed Form Equations)
3: If ( SP<UB), Update the UP and the best point
4: Add the OA inequalities and solve the MP
(No need to solve MP to optimality, Fletcher and Leyffer; 1994)
5: Update the LB, h=h+1, go to step 1
6:Report the Solutions and Algorithm Stops
18. Networks Number of Links(A) Number of Nodes (N)
SiouxFalls 76 24
Anaheim 914 416
Barcelona 2522 1020
Chicago Sketch 2250 933
Experiments Set up
OA-is coded in GAMS-C++/CPLEX
MILP MP is solved via CPLEX solver
MP is solved within 1% optimality gap
Covariance matrix is randomly generated
20. 0
20
40
60
80
100
120
140
160
180
0.5 0.6 0.8 1 1.5 2
RunningTime(Sec)
Master Problem Relative Optimality Gap(%)
Anaheim Sioux Falls Chicago Sketch Barcelona
Solution time is stable w.r.t. MP optimality gap
21. 0
50
100
150
200
250
0 0.2 0.4 0.6 0.8 1
RunningTime(Sec)
Weight Associated with Mean Tavel Cost
Sioux Falls Anaheim Chicago Sketch Barcelona
Increase in weight on STD term increases the
solution time for larger networks
23. General Reliability Measures
[Shahabi, Unnikrishnan, and Boyles, 2014, 2015] show
application of OA to other forms of convex reliability metrics
Joint Inventory Location Problems
[Shahabi, Unnikrishnan, and Boyles, 2014] show that OA is
efficient in solving multi-echelon joint inventory location
problems with correlations in retailer demand
24. •A general efficient methodology for binary convex
mixed integer programs was presented
•Given the convexity assumption the method
guarantees the global optimality
•The sub-problem is analytically solved
•Minimizes the impact of large correlation matrices
and lead to savings in computational time
Notes de l'éditeur
There are different ways to characterize reliability and risk taking behavior…One such metric is mean-std dev..this ppt is mainly going to focus on mean+std dev shortest path..
CQP computationally tractable due to their special Structure
Can be solved by polynomial time interior point algorithms
Solvers like CPLEX and MOSEK offering this capability