Constraint-based Approaches for Balancing Bike Sharing Systems

Transcription

Constraint-based Approaches for Balancing Bike Sharing Systems
Constraint-based Approaches for Balancing
Bike Sharing Systems
Luca di Gaspero1 , Andrea Rendl2 and Tommaso Urli1
DIEGM, University of Udine,
Via Delle Scienze, 206 - 33100 Udine, Italy
{luca.digaspero|tommaso.urli}@uniud.it
Dynamic Transportation Systems, Mobility Department,
Austrian Institute of Technology
Giefinggasse 2, 1210 Vienna, Austria
[email protected]
in cooperation with TU Vienna - Algorithms and Data Structures Group
funded by the Austrian Federal Ministry of Transport, Innovation and Technology (BMVIT)
CP-2013, September 17, 2013
Bike Sharing Systems
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
2 / 26
Unbalanced Bike Stations
stations are located all over the city
load of bikes varies strongly between stations
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
3 / 26
Balancing Bike Stations
station load becomes unbalanced over time
bikes need to be re-distributed
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
4 / 26
Applied Research Project
Project Partners
•
•
•
•
Austrian Institute of Technology (AIT)
TU Vienna, ADS Group
Citybike Wien/Vienna
Energie- und Umweltagentur NÖ (environemental agency)
cooperation with Luca di Gaspero and Tommaso Urli from
the University of Udine, Italy
Project aim: develop a decision support system for
balancing bike sharing systems
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
5 / 26
Applied Research Project
Project Partners
•
•
•
•
Austrian Institute of Technology (AIT)
TU Vienna, ADS Group
Citybike Wien/Vienna
Energie- und Umweltagentur NÖ (environemental agency)
cooperation with Luca di Gaspero and Tommaso Urli from
the University of Udine, Italy
Project aim: develop a decision support system for
balancing bike sharing systems
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
5 / 26
Balancing Bike Sharing Systems Problem (BBSS)
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
6 / 26
Balancing Bike Sharing Systems Problem (BBSS)
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
7 / 26
Static BBSS: Problem Parameters
bike stations S = {1, . . . , S}
• capacity Cs
• load bs
• target load ts (demand)
fleet of vehicles V = {1, . . . , V }
• capacity cv
• depot D
• time budget t̂
travel time matrix travelTimeu,v with u, v ∈ S ∪ D
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
8 / 26
Static BBSS
Find a tour with loading instructions for each vehicle,
considering:
•
•
•
•
•
vehicle capacity
station capacities
time budget
vehicles start and end their tours empty at the depot
a vehicle does not visit a station more than once
objective: minimal deviation from the target loads and
minimal travel time
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
9 / 26
Static BBSS
Find a tour with loading instructions for each vehicle,
considering:
•
•
•
•
•
vehicle capacity
station capacities
time budget
vehicles start and end their tours empty at the depot
a vehicle does not visit a station more than once
objective: minimal deviation from the target loads and
minimal travel time
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
9 / 26
1
CP Models for BBSS
Routing Model
Step Model
2
Large Neighbourhood Search (LNS)
3
Experimental Results
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
10 / 26
Routing Model
Based on Vehicle Routing Problem (VRP) model:
extended graph: start and end depot for each vehicle
successor variables for each station
dummy vehicle visits unvisited stations
service and load variables for loading instructions
vehicle and time variables
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
11 / 26
Routing Model
Based on Vehicle Routing Problem (VRP) model:
extended graph: start and end depot for each vehicle
successor variables for each station
dummy vehicle visits unvisited stations
service and load variables for loading instructions
vehicle and time variables
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
11 / 26
Routing Model
2 vehicles and 5 stations
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
12 / 26
Routing Model: Search Strategy
incrementally construct the tours vehicle by vehicle
successor variables:
• variable selection: the successor of the last variable
• value selection: according to the stations’ utility
• if time budget is consumed for current vehicle, the
successor is set to the next vehicle start depot
service variables:
after setting each successor variable, we search on its
respective service variable
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
13 / 26
Routing Model: Search Strategy
incrementally construct the tours vehicle by vehicle
successor variables:
• variable selection: the successor of the last variable
• value selection: according to the stations’ utility
• if time budget is consumed for current vehicle, the
successor is set to the next vehicle start depot
service variables:
after setting each successor variable, we search on its
respective service variable
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
13 / 26
Routing Model: Search Strategy
incrementally construct the tours vehicle by vehicle
successor variables:
• variable selection: the successor of the last variable
• value selection: according to the stations’ utility
• if time budget is consumed for current vehicle, the
successor is set to the next vehicle start depot
service variables:
after setting each successor variable, we search on its
respective service variable
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
13 / 26
Step Model
Based on an AI-planning perspective
route for each vehicle of maximal K stops, starting and
ending at the depot
estimated upper bound for K
service, load and time variables
advantage: direct representation each vehicle’s route
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
14 / 26
Step Model
2 vehicles and 5 stations
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
15 / 26
Step Model: Search Strategy
construct tour after tour, searching on route and service
variables
route variables: static variable selection, selecing the
station with highest utility (deviation from target)
service variables: after setting each route variable, search
on the service with dynamic max-value selection
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
16 / 26
Large Neighbourhood Search (LNS)
given a valid solution
destroy-step: release parts of the solution (large
neighbourhood), fix the remaining variables
repair-step: solve problem with an exact approach to
optimality
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
17 / 26
Large Neighbourhood Search (LNS): destroy
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
18 / 26
Large Neighbourhood Search (LNS): repair
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
19 / 26
LNS Parameters (1/2)
initial solution: first solution from tree search
destruction rate d increases if no improvement can be
made and is reset otherwise
destroy step
Routing model select d · |Ri | stations from each tour Ri for
destruction, and reset successor, service and
vehicle variables
P
Step model select d · i |Ri | visited stations and reset
route and service variables
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
20 / 26
LNS Parameters (2/2)
repair step: Branch & Bound with time limit proportional to
the number of free variables
acceptance: the repaired solution is accepted if it
improves the current best solution
restarts: after C iterations with no improvement, a new
initial solution is generated and search is restarted
stopping criterion: timeout
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
21 / 26
Experimental Setup
adapted real world instances derived from historical data
LNS parameters are tuned by F-Race (confidence level
0.95 over 150 instances)
we provided our source code and instances to the
recomputation initiative (see Tutorial by Ian Gent and Lars
Kotthoff on Wednesday, 13:30)
CP Solver: Gecode 3.7.3
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
22 / 26
Routing versus Step Model: CP and LNS
red: routing model, blue: step model
dotted: pure CP, line: average LNS
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
23 / 26
LNS versus State-of-the-Art
state-of-the-art: VNS approach (Raidl et al, 2013)
our approach is competitive but does not beat the
state-of-the-art
advantage of our approach: easily extendable to similar
problem setups
detailed results are in our paper
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
24 / 26
Conclusions & Future Work
What to remember form this talk:
•
•
•
•
Nice problem: BBSS
routing and step model
Large Neighbourhood Search (LNS)
CP approach is competitve with state-of-the-art
Future Work: dynamic multi-day BBSS
L. di Gaspero, A. Rendl, T. Urli |
17.9.2013
25 / 26
Constraint-based Approaches for Balancing
Bike Sharing Systems
Luca di Gaspero1 , Andrea Rendl2 and Tommaso Urli1
DIEGM, University of Udine,
Via Delle Scienze, 206 - 33100 Udine, Italy
{luca.digaspero|tommaso.urli}@uniud.it
Dynamic Transportation Systems, Mobility Department,
Austrian Institute of Technology
Giefinggasse 2, 1210 Vienna, Austria
[email protected]
in cooperation with TU Vienna - Algorithms and Data Structures Group
funded by the Austrian Federal Ministry of Transport, Innovation and Technology (BMVIT)
CP-2013, September 17, 2013