robust optimization with recovery: application to...
TRANSCRIPT
![Page 1: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/1.jpg)
1
Robust Optimization with Recovery: Application to Shortest Paths and
Airline Scheduling
Niklaus EggenbergDr. Matteo Salani and Prof. Michel Bierlaire
Funded by SNF
![Page 2: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/2.jpg)
2
Planning the Recovery
Motivation
Developed Recovery Algorithm for disrupted airline schedules
Want to consider the possibility of recovery during scheduling phase
![Page 3: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/3.jpg)
3
Index
Index
Optimization under Uncertainty
Approach classification
Proactive worst case approach including recovery
decisions
Illustration on the SPPID
Application to Airline Scheduling
Future work and conclusions
![Page 4: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/4.jpg)
4
Optimization under Uncertainty(for a minimization problem)
Optimization under uncertainty
Characterization of uncertainty set U• any knowledge?• probabilistic measure?• consider extreme cases?
Proactive vs Reactive approach
![Page 5: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/5.jpg)
5
Existing Methods’ Classification
Classification
Proactive Methods Reactive Methods
Probabilistic UStochastic
Programming
StochasticProgramming with
Recourse
Non Probabilistic U Robust Optimization On-line Algorithms
![Page 6: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/6.jpg)
6
Some References• On-Line Algorithms: Albers, 2003
• Stochastic Optimization: Kall and Wallace, 1994; Wallace and Ziemba, 2005.
• Stochastic with Recourse: Kall and Wallace, 1994; Polychronopoulos and Tsitsiklis , 1996; Provan, 2003.
• Robust Optimization: Soyster, 1973; Ben-Tal and Nemirowski, 2001 (book); Bertsimas and Sim, 2004.
Classification - References
![Page 7: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/7.jpg)
7
On-Line Optimization
Classification (5)
Advantages Disadvantages
• react in real time
• easy computation (decision strategy)
• optimality ratio (a posteriori)
• no information on bounds (only ratio)
• reaction dictated by nature
• difficult to measure performance a
priori
![Page 8: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/8.jpg)
8
Stochastic Programming
Classification (2)
Advantages Disadvantages
• lowest expected cost
• low probabilities of high cost
• exploits a priori information
• difficult to measure probabilities
• estimation of all scenarios
• only predictor (depends on scenario)
• not efficient for few implementations
![Page 9: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/9.jpg)
9
Stochastic Programming with Recourse
Classification (3)
Advantages Disadvantages
• includes cost of recourse decision
• adapts to information in real time
• exploits partial information
• difficult to measure probabilities
• estimation of all scenarios
• predictor on partially revealed scenario
• guided by information revelation
![Page 10: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/10.jpg)
10
Robust Optimization
Classification (4)
Advantages Disadvantages
• solution always feasible
• evaluation on worst scenarios only
• get cost bound
• characterize worst scenarios
• what if a good scenario occurs?
![Page 11: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/11.jpg)
11
Definitions
Notation
Problem P
• set of feasible solutions S
• uncertainty set U
• realization or scenario u U
• cost of solution s under scenario u: cu(s)
![Page 12: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/12.jpg)
12
Recoverable Approach
Proactive Worst Case Algorithm with Recovery
• Proactive: want a deterministic solution not dictated by scenario revealing
• Worst + Best: consider both best and worst scenarios (potential)
• Include On-line Decisions (Recovery): consider recovery costs for unfeasible solutions
![Page 13: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/13.jpg)
13
Example:
Example: SPPID
Shortest Path Problem with Interval Data (SPPID):
• Oriented graph G = (V,A)
• Unique source s and unique sink t
• Random arc costs cij [lij, uij] (uij = possible)
• Scenario: set of cost realization for every arc cost
• arc cost revealed when source node is reached
![Page 14: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/14.jpg)
14Example: SPPID (2)
On-Line Strategy:
• at node i take outgoing arc with lowest cost
Recovery:
• when reach dead end, take last arc backwards at cost uij
• remove arc from V
Probabilistic cases:
• symmetric independent distributions
![Page 15: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/15.jpg)
15Example: SPPID (3)
{1, }
[3,5]
![Page 16: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/16.jpg)
16Example: SPPID (4)
{1, }• Stochastic: path {s,e,f,t}, mean cost 27
• Robust: path {s,d,t}, worst cost 33
• On-line: either {s,a,b,t} or {s,e,f,t}
• Stochastic with recourse: either {s,e,f,t} or {s,a,b,t}
![Page 17: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/17.jpg)
17Example: SPPID (5)
{1, }• Recoverable: path {s,a,b,t}
Path Best Worst Potential Mean
{s,a,b,t} 11 42 53
{s,a,c,t} 22 34 56 28
{s,d,t} 28 33 61 30.5
{s,e,f,t} 17 37 54 27
![Page 18: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/18.jpg)
18Example: SPPID (6)
Method Path Cost
On-Line {s,e,f,t} 17
Stochastic with Recourse
{s,e,f,t} 17
Stochastic {s,e,f,t} 17
Robust {s,d,t} 28
Recoverable{s,a,b,t}
Or{s,a,b,a,c,t}
11or 28
{1, }
3 410
13 15
131
3
2
![Page 19: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/19.jpg)
19Example: SPPID (7)
Method Path Cost
On-Line{s,a,b,t} or {s,a,b,a,c,t}
1742
Stochastic with Recourse
{s,a,b,t} or {s,a,b,a,c,t}
1742
Stochastic {s,e,f,t} 37
Robust {s,d,t} 33
Recoverable{s,a,b,t} or{s,a,b,a,c,t}
17 42
{1, }
12 814
16 17
1715
5
4
![Page 20: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/20.jpg)
20Example: SPPID (8)
Method Path Cost
On-Line {s,e,f,t} 26
Stochastic with Recourse
{s,a,b,t} or {s,a,b,a,c,t}
1538
Stochastic {s,e,f,t} 26
Robust {s,d,t} 29
Recoverable{s,a,b,t} or{s,a,b,a,c,t}
15 38
{1, }
11 713
14 15
1310
3
3
![Page 21: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/21.jpg)
21Example: SPPID (9)
Method Path Cost
On-Line {s,e,f,t} 25
Stochastic with Recourse
{s,e,f,t} 25
Stochastic {s,e,f,t} 25
Robust {s,d,t} 29
Recoverable{s,a,b,t} or{s,a,b,a,c,t}
11 28
{1, }
8 410
13 16
173
5
2
![Page 22: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/22.jpg)
22Example: SPPID (10)
Method Path Cost
On-Line Scenario dependent 26.875
Stochastic with Recourse
Scenario dependent 26.375
Stochastic {s,e,f,t} 26.250
Robust {s,d,t} 29.875
Recoverable{s,a,b,t} or{s,a,b,a,c,t}
24.250
Average costs over 8 different scenarios
![Page 23: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/23.jpg)
23Example: Summary
• Recoverable solution both highest and lowest cost
• Robust and Stochastic reject possibly unfeasible paths
• Reactive dependent on scenario
REMARK
Path {s,a,b,t} has least expected cost (24.5) if consider stochastic with recourse in a proactive way
![Page 24: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/24.jpg)
24Recoverable Algorithm for Airlines
Recoverable Schedule for Airlines
Difficulties:
• evaluation of feasibility
• recovery problem NP-hard
• characterization of worst scenario impossible
• evaluation on whole U impossible
![Page 25: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/25.jpg)
25Recoverable Algorithm for Airlines (2)
Research Directions:
• Scenario Sampling
• Uncertainty set structure
• Recovery costs estimations
• Implicit measures and multi-objective optimization
• Explicit measures in recovery algorithm
![Page 26: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/26.jpg)
26Future Work
Future Work:
• Develop a specific formalism
• Explore more deeply possible directions
• Develop an efficient algorithm
• Benchmark with real problems
![Page 27: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/27.jpg)
27Conclusions
Conclusions:
• Proactive best-worst case framework including information on unfeasible solutions
• Few scenario estimations, no probabilistic measures
• Provides recovery plan for each scenario
• Trade-off between proactive and reactive (planning
the recovery)
![Page 28: Robust Optimization with Recovery: Application to …transp-or.epfl.ch/documents/talks/NEgSTRC07.pdf · 1 Robust Optimization with Recovery: Application to Shortest Paths and Airline](https://reader036.vdocuments.us/reader036/viewer/2022081605/5b5e980f7f8b9a6d448c8d97/html5/thumbnails/28.jpg)
28
THANKS for your attention!
Any Questions?