Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Determination of robust solutions for the DARP with variations in
transportation time
Maxime Chassaing, Christophe Duhamel, Gérard Fleury and Philippe Lacomme
LIMOS, Université Blaise Pascal
France, Clermont-Ferrand
Contact : [email protected]
Stochastic DARP Problem definition
• Dial-a-Ride Problem (DARP)
• Stochastic travel times
Solution
• Associated driver policy
• Robustness
Evaluation in the stochastic context
• Indirect robustness measure
• Simulation / Analytic evaluation
Resolution methods
• Iterative: ELS / Bi-criteria: NSGA-2
Conclusion
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
1
29
/06
/20
16
Stochastic problems and vehicle routing problems
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
2
Type 1 Variation of collected volume • Stochastic CARP : (Fleury et al., 2008)
Type 2 Variation of clients to be served • (Heilporn et al., 2011)
Type 3 Variation of travel times. • breakdowns, • route hazards, • congestion, traffic flow Review : (Gendreau et al., 2015) Stochastic Shortest Path (DSSPP) (Pattanamekar et al., 2003) VRP with Time Windows (Tas et al., 2013)
Publications
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
3
Dial-a-ride Problem (DARP)
• Instances: (Cordeau and Laporte, 2003) (Ropke and Laporte, 2007)
• Methods VNS (Parragh and Schmid, 2013) ALNS (Masson et al., 2014) DA (Braekers et al., 2014)
Stochastic DARP (dynamic)
• On client’s demand
• (Heilporn et al., 2011)
• On travel time
• (Xiang et al., 2008)
• (Schilde et al., 2014)
event (crash) Influence area
Dial-a-Ride Problem (DARP)
Definition
• Homogeneous fleet of 𝑴 vehicles (𝑀 ≥ 1) :
• capacity : 𝐶𝑚𝑎𝑥
• 1 depot node
• 𝑵 clients (𝑁 ≥ 1) :
• 𝑅𝑖 request ( origin 𝐼+ destination 𝐼−)
Constraints
• [𝑒𝑖𝑚𝑖𝑛, 𝑙𝑖
𝑚𝑎𝑥] time windows : TW
• Total duration : 𝑻𝑹𝑻max
• Maximal riding time : 𝑻𝑫max
OBJECTIVE: find a solution to transport each client and
minimizing the total distance
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
4
0
1+
2-
3+
1-
3-
1
1
2
2+
20,40
15,35
35,50
30,60
80,90
10,20
85 60
50
35
2515
Solution of DARP instances
• solution = set of trips
• trip is
vertex order :
depot , 1+, 1-, 2+, 3+, 2-, 3-, depot
set of variables for each vertex
• 𝑨𝒊 : arrival time on i
• 𝑩𝒊 : beginning of service at i
• 𝑾𝒊 : waiting time on i
• 𝑫𝒊 : departure time from i
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
5
(Cordeau et Laporte, 2003)
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
6
Di departure time from vi to vj
i
j
ej ljTime windows on vj
Aj arrival date on vj
Ai arrival Date on vi
Bi beginning of service
Wi waiting time
di service duration
Time windows on vi
ei li
time
no
de
s
Dj
Bj
tij Wj
(Cordeau et Laporte, 2003)
Deterministic case : 𝑩𝒊 (known) ⟹ 𝑫𝒊(known) ⟹ 𝑨𝒊+𝟏 (known) Evaluation : (Cordeau et al. 2003), (Firat and Woeginger, 2011)
Illustration of trip’s variables (specific to the DARP)
Di departure time from vi to vj
i
j
ej ljTime windows on vj
Aj arrival date on vj
Ai arrival Date on vi
Bi beginning of service
Wi waiting time
di service duration
Time windows on vi
ei li
time
Dj
Bj
Tij(ω’) Wj
no
de
s
Tij(ω)
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
7
For the stochastic case : 𝑩𝒊 (known) ⟹ 𝑫𝒊(known) ⇏ 𝑨𝒊+𝟏 (unknown)
Stochastic travel times (implies choices on policy of drivers)
Illustration of trip’s variables
Main model: normal distribution
Tij~𝑁(𝑡𝑖𝑗 , 𝜎𝑖𝑗) with 𝜎𝑖𝑗= 𝑡𝑖𝑗/10
• example : 𝑡𝑖𝑗 = 10 min
• 𝑃 (9 𝑚𝑖𝑛 ≤ 𝑇𝑖𝑗 ≤ 11 𝑚𝑖𝑛) ≅ 70%
Other model: gamma distribution (Tas et al., 2013) shifted gamma distribution
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
8
Travel time model used
expected value = 𝑡𝑖𝑗
Standard deviation = 𝑡𝑖𝑗
𝛾
Disadvantages :
- for a given 𝛾 expected value and variance are linked
𝑡𝑖𝑗 0
𝛾=4
68,2%
15,9% 15,9%
𝑡𝑖𝑗 𝑡𝑖𝑗 -𝜎𝑖𝑗 𝑡𝑖𝑗 +𝜎𝑖𝑗
0
expected value variance
𝑡𝑖𝑗 0
𝛾=1
𝑡𝑖𝑗 0
𝛾=16
• Using the expected values
• 𝑨𝒊 ; 𝑩𝒊 ; 𝑾𝒊 ; 𝑫𝒊 can be computed
But for realizations: 𝝎 and 𝝎′ travel time became 𝐓𝐢𝐣(𝝎) and 𝐓𝐢𝐣(𝝎′)
• So, become random values:
𝑨𝒊(𝝎) ; 𝑩𝒊(𝝎) ; 𝑾𝒊(𝝎) ; 𝑫𝒊(𝝎)
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
9
si
si-1
li-1ei-1
Di-1
𝐴𝑖(𝜔)
Bi-1
𝐵𝑖
𝐴𝑖(𝜔′)
𝑇𝑖−1,𝑖(𝜔′)
Ai-1 Wi-1 di-1
𝑇𝑖−1,𝑖(𝜔)
𝐴𝑖 𝑊𝑖
theoretical value
Consequences of stochastic travel time
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
10
Driver’s policies
Because 𝐴𝑖 𝜔 ≠ 𝐴𝑖 the driver have different choices:
• 𝑖𝑓 𝐴𝑖 𝜔 > 𝐴𝑖 then ???
• 𝑖𝑓 𝐴𝑖 𝜔 < 𝐴𝑖 then ???
Different driver’s policies can be defined: 1 « if the driver is early then he waits until 𝐵𝑖 »
2 « the driver has to respect the waiting time 𝑊𝑖 »
3 …
si
si-1
li-1ei-1
Di-1
𝐴𝑖(𝜔)
Bi-1
𝐵𝑖
Ai-1 Wi-1 di-1
𝑇𝑖−1,𝑖(𝜔)
𝐴𝑖
theoretical value
realization value
𝑊𝑖
A solution of DARP
• order of vertices
• Defined dates 𝑨𝒊, 𝑩𝒊, 𝑾𝒊, 𝑫𝒊
Robustness 𝑷(𝑺) : probability that the solution satisfies each constraint
𝑃(𝑆) = 𝑃 𝑡𝑖𝑖
with 𝑃 𝑡𝑖 probability that a trip is valid
Driver’s policy : meet 𝑩𝒊, 𝐞𝐧𝐬𝐮𝐫𝐞 𝑾𝒊, …
New criterion : robustness
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
11
S2
S1
S3 S4
x
Vertex order ‘s space
2 – Robustness estimator
S2 ( 180 , 0.5)
S1 ( 180 , 0.1)
S3 ( 180 , 0.7)
S4 ( 180 , 0.9)
2
Solutions space (Cost known)
3
1 - Evaluation: compute Bi
Keep the solutionwith the highest
robustness
1 S -> (Cost(S), Robustness = F(S,n))
PROPOSITION
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
12 Evaluation: compute Bi
maximizing a criterion of robustness ρ*
1'Directly used
ρ* as criterion of robustness
2'
0
1+
2-
3+
1-
3-
1
1
2
2+
20,40
15,35
35,50
30,60
80,90
10,20
85 60
50
35
2515
0
1+
2-
3+
1-
3-
1
1
2
2+
20,40
15,35
35,50
30,60
80,90
10,20
80 55
45
30
2010
1- Evaluation for the SDARP
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
si
si-1
Di-1 Bi-1
𝐴𝑖 𝐴𝑖𝜔
di-1
𝑇𝑖−1,𝑖(𝜔)
si
si-1
Di-1 Bi-1
𝐴𝑖 𝐴𝑖𝜔
di-1
𝑇𝑖−1,𝑖(𝜔) 13
Evaluation for the DARP is, for a fixed order of clients, compute Bi on vertices which respect constraints
This schedule affects the robustness
Definition of valid constraint Deterministic case
• Constraints : 𝐵𝑖 ≤ 𝐶𝑠𝑡
𝐵𝑖 ≥ 𝐶𝑠𝑡
𝐵𝑖 − 𝐵𝑗 ≤ 𝐶𝑠𝑡
Stochastic case
𝐵𝑖(𝜔) depends on realizations of travel time
- for some 𝜔 the constraint is valid
- for some 𝜔 ′ the constraint is not valid
PROPOSITION : In the model a constraint is consider valid to the order ρ if:
𝑷{ 𝐵𝑖(𝜔) ≤ 𝐶𝑠𝑡 } ≥ ρ ⇔ 𝐵𝑖 +𝑐(ρ)𝜎𝑖 ≤ 𝐶𝑠𝑡
𝑷{ 𝐵𝑖 𝜔 ≥ 𝐶𝑠𝑡 } ≥ ρ ⇔ 𝐵𝑖 +𝑐(ρ)𝜎𝑖≥ 𝐶𝑠𝑡
𝑷{𝐵𝑖(𝜔) − 𝐵𝑗(𝜔) ≥ 𝐶𝑠𝑡 } ≥ ρ ⇔ 𝐵𝑖 − 𝐵𝑗 + 𝑐 ρ (𝜎𝑖 + 𝜎𝑗) ≤ 𝐶𝑠𝑡
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
𝐶𝑠𝑡
𝐵𝑖 j
𝑐(𝜌)𝜎𝑖
stochastic problem :
𝐶𝑠𝑡
𝐵𝑖 j
deterministic problem
𝐵𝑖 ≤ 𝐶𝑠𝑡
14
PROPOSITION: compute 𝐵𝑖 for a fixed ρ
Using the definition of valid constraint to the order ρ
Evaluate a trip is possible if is known :
- the driver’s policy
- the ρ value
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
15
evaluation proposed by (Firat and Woeginger , 2011) using the new constraints
A+ B+ C+0 C-0 00
B- A-00
CST CST CSTCST
CST
-CST-CST -CST -CST -CST
CST
-CST
00
-CST
-CST
-CST
i=0 i=1 i=2 i=4i=3 i=5 i=6
CST
Time Windows min constraints Time Windows max constraints Riding time max and total duration max constraints
PROPOSITION: to improve the solution robustness
Maximize a criterion of robustness
Objective : For each trip
find the maximal value of ρ possible: ρmax 𝑡
• Dichotomy to find ρmax 𝑡
using ρmax 𝑡 of each trip a criterion associated to the robustness
of a solution have been created:
ρ* = ρmax 𝑡𝑡
ρ* will by used in the metaheuristic to replace 𝑃(𝑆)
(Warning) ρ* ≠ 𝑃(𝑆)
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
16
For a driver policy and affected dates on vertices
• The robustness of a solution can be compute :
• analytically 𝑷(𝒔)
• by simulation 𝑭 𝒔, 𝒏 an estimator of 𝑃(𝑆)
with 𝑛 number of replications
2 - Robustness estimator
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
17
100%
Simulation n (replication)
𝐹(𝑠, 𝑛)
𝑃(𝑠)
S2
S1
S3 S4
x
Vertex order ‘s space
2 – Robustness estimator
S2 ( 180 , 0.5)
S1 ( 180 , 0.1)
S3 ( 180 , 0.7)
S4 ( 180 , 0.9)
2
Solutions space (Cost known)
3
1 - Evaluation: compute Bi
Keep the solution with the highest robustness
1 S -> (Cost(S), Robustness = F(S,n))
Evaluation: compute Bimaximizing a criterion of
robustness ρ*
1'
Directly used ρ* as
criterion of robustness
2'
On the 20 instances proposed by (Cordeau et Laporte, 2003)
BKS (Best known solution): (Braekers et al., 2014) (Parragh et Schmid, 2013)
Results for normal distribution
H1 : driver have to met Bi
𝑩𝒊 theoretical beginning of service
Comparison between ρ* and 𝐹 𝑠, 𝑛
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
18
name m n c(BKS) ρ𝐻1∗ F𝐻1
(BKS, n)
pr01 3 24 190.02 74.8% 86.4%
pr02 5 48 301.34 76.8% 85.4%
pr03 7 72 532.00 4.8% 0.1%
pr04 9 96 570.25 14.3% 0.3%
pr05 11 120 626.93 13.1% 0.7%
pr06 13 144 785.26 3.9% 0.0%
pr07 4 36 291.71 21.9% 14.7%
pr08 6 72 487.84 8.9% 0.2%
pr09 8 108 658.31 9.4% 0.4%
pr10 10 144 851.82 1.3% 0.0%
pr11 3 24 164.46 63.2% 61.0%
pr12 5 48 295.66 32.0% 5.8%
pr13 7 72 484.83 32.5% 16.5%
pr14 9 96 529.33 59.3% 64.6%
pr15 11 120 577.29 37.9% 72.0%
pr16 13 144 730.67 16.2% 11.1%
pr17 4 36 248.21 26.8% 1.6%
pr18 6 72 458.73 62.2% 72.9%
pr19 8 108 593.49 5.8% 0.0%
pr20 10 144 785.68 3.8% 0.0%
avg. 28.4% 24.7%
value
instances
Generate robust solutions
Two methods :
1 – mono-criterion with
1𝑠𝑡 : 𝑃(𝑆) (solution robustness) ρ*(criteria associated)
2𝑛𝑑 : 𝐶(𝑆) (solution cost)
Evolutionary local search (ELS) (Wolf and Merz, 2007)
2 – multi-criteria : a solution label (ρ∗(𝑆), 𝐶(𝑆))
NSGA-II algorithm
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
19
C(S)
ρHx(S)
optimal Pareto front
10
robustness
criterion ρ
distance
dominated solutions
Results BKS vs BFS
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
20
𝑩𝑲𝑺 𝑩𝑭𝑺 normal distribution 𝑩𝑭𝑺 Shifted gamma
distribution 𝜸 = 𝟒
name m n c(BKS) F𝐻1(BKS, n) c(BFS) gap F 𝐻1(BFS, n) time
(min) c(BFS) gap F 𝐻1(BFS, n)
time
(min)
pr01 3 24 190.02 86.4% 200.48 5.50% 100.0% 0.25 210.94 11.01% 99.9% 0.25
pr02 5 48 301.34 85.4% 307.11 1.92% 100.0% 1.25 317.00 5.20% 100.0% 1.25
pr03 7 72 532.00 0.1% 587.96 10.52% 100.0% 2.30 628.59 18.16% 99.7% 2.30
pr04 9 96 570.25 0.3% 643.75 12.89% 100.0% 7.37 663.68 16.38% 99.3% 7.37
pr05 11 120 626.93 0.7% 744.92 18.82% 100.0% 12.07 752.20 19.98% 99.8% 12.07
pr06 13 144 785.26 0.0% 979.78 24.77% 100.0% 21.92 993.02 26.46% 99.1% 21.92
pr07 4 36 291.71 14.7% 306.69 5.13% 100.0% 0.47 320.29 9.80% 100.0% 0.47
pr08 6 72 487.84 0.2% 573.65 17.59% 100.0% 2.68 607.81 24.59% 99.3% 2.68
pr09 8 108 658.31 0.4% 774.53 17.65% 100.0% 11.25 790.74 20.12% 93.8% 11.25
pr10 10 144 851.82 0.0% 1049.23 23.18% 99.5% 21.33 1052.57 23.57% 76.6% 21.33
pr11 3 24 164.46 61.0% 168.81 2.64% 100.0% 0.28 178.97 8.82% 100.0% 0.28
pr12 5 48 295.66 5.8% 310.66 5.07% 100.0% 1.37 334.41 13.10% 100.0% 1.37
pr13 7 72 484.83 16.5% 527.10 8.72% 100.0% 3.70 605.94 24.98% 100.0% 3.70
pr14 9 96 529.33 64.6% 634.16 19.80% 100.0% 10.20 638.52 20.63% 100.0% 10.20
pr15 11 120 577.29 72.0% 634.08 9.84% 100.0% 19.93 701.94 21.59% 99.8% 19.93
pr16 13 144 730.67 11.1% 891.86 22.06% 100.0% 32.32 934.23 27.86% 99.9% 32.32
pr17 4 36 248.21 1.6% 266.83 7.50% 100.0% 0.58 294.83 18.78% 100.0% 0.58
pr18 6 72 458.73 72.9% 494.54 7.81% 100.0% 4.32 530.68 15.68% 100.0% 4.32
pr19 8 108 593.49 0.0% 699.70 17.89% 100.0% 12.43 712.33 20.02% 99.9% 12.43
pr20 10 144 785.68 0.0% 978.61 24.56% 100.0% 31.45 972.21 23.74% 99.8% 31.45
avg.
24.7% 13.19% ~100% 9.87 18.52% 98,4% 9.87
BKS: Best Known Solution for the DARP determinist (Braekers et al., 2014) (Parragh et Schmid, 2013) (Cordeau et Laporte, 2003)
BFS: Best Found Solution obtain with the ELS
Results with NSGA-2 ( Extreme solutions of the front )
column with avg. of 5 runs : 𝑛𝑏 ; 𝑔𝑎𝑝 ; 𝐹𝐻1(𝑆, 𝑛)
column with best results of 5 runs : 𝑔𝑎𝑝* ; 𝐹𝐻1(𝑆, 𝑛) *
MIM
20
16
: 8
th IF
AC
co
nfe
ren
ce
29
/06
/20
16
name Solution with minimal cost Solution with maximal robustness temps
𝒏 𝒈𝒂𝒑 𝒈𝒂𝒑∗ 𝐅𝐇𝟏(𝐬, 𝐧) 𝑭𝑯𝟏(𝑺, 𝒏) * 𝒈𝒂𝒑 𝒈𝒂𝒑∗ 𝐅𝐇𝟏(𝐬, 𝐧) 𝑭𝑯𝟏(𝑺, 𝒏)
* min
pr01 14.6 0.08% 0.00% 87.3% 85.4% 3.87% 4.31% 100.0% 100% 0.25
pr02 20.6 2.18% 0.86% 79.2% 90.3% 4.17% 5.51% 100.0% 100% 1.25
pr03 42.2 4.13% 2.84% 55.2% 86.8% 12.40% 11.03% 100.0% 100% 2.30
pr04 43.0 5.85% 3.18% 55.4% 32.9% 9.95% 12.38% 100.0% 100% 7.37
pr05 49.4 6.88% 5.16% 31.4% 38.6% 10.61% 12.21% 100.0% 100% 12.07
pr06 73.4 6.23% 4.24% 12.6% 0.3% 10.84% 10.08% 100.0% 100% 21.92
pr07 13.4 2.11% 1.21% 63.6% 81.5% 5.12% 5.79% 100.0% 100% 0.47
pr08 49.6 5.35% 3.61% 31.5% 37.3% 18.87% 20.09% 100.0% 100% 2.68
pr09 84.0 8.06% 5.90% 20.6% 5.8% 18.01% 13.18% 100.0% 100% 11.25
pr10 89.0 8.17% 7.02% 2.5% 2.2% 17.51% 20.80% 95.5% 100% 21.33
pr11 7.6 2.19% 0.25% 78.3% 75.1% 3.71% 4.52% 100.0% 100% 0.28
pr12 20.4 3.77% 2.72% 57.8% 26.3% 5.98% 5.89% 100.0% 100% 1.37
pr13 38.0 7.39% 6.22% 64.3% 43.3% 13.20% 12.23% 100.0% 100% 3.70
pr14 29.8 7.37% 6.66% 52.1% 95.3% 11.58% 12.92% 100.0% 100% 10.20
pr15 30.0 5.77% 2.09% 23.6% 48.4% 8.28% 9.33% 96.4% 100% 19.93
pr16 72.8 7.48% 6.47% 22.5% 38.7% 11.37% 9.94% 100.0% 100% 32.32
pr17 19.2 4.11% 2.04% 31.0% 5.7% 8.86% 10.80% 100.0% 100% 0.58
pr18 29.0 5.25% 3.11% 47.3% 78.1% 11.10% 12.17% 100.0% 100% 4.32
pr19 61.0 8.16% 6.80% 15.7% 1.2% 15.92% 22.34% 100.0% 100% 12.43
pr20 69.6 8.73% 7.59% 12.3% 27.1% 14.47% 19.04% 99.9% 100% 31.45
avg. 42.8 5.46% 3.90% 42.2% 45.0% 10.79% 11.73% 99.6% 100% 9.87
21
Conclusion Proposition
1 - a robustness criterion (usable in a metaheuristic) 2 - an evaluation method to obtain robust schedules (according to a driver’s policy)
Methods General framework to optimize the robustness
Results obtained For classic literature DARP instances. 1 - the BKSs for the determinist DARP are not robust 2 - the solutions we computed provide a « reasonable » cost Detailed results are available online : http://fc.isima.fr/~chassain/SDARP/SDARP.php
29
/06
/20
16
M
IM2
01
6 :
8th
IFA
C c
on
fere
nce
22