Page 1
Michael Szell [email protected]
Paolo SantiGiovanni RestaStanislav SobolevskyCarlo RattiSteven Strogatz (Cornell)
Benedikt GroßJoey LeeEric BaczukCarlo RattiAndi Weiß (47Nord)Stefan Landsbeck (47Nord)
Research Visualization & Explorer
hubcabTaxi-sharing in New York City: A network-based approach
Page 2
Large GPS data sets on taxi movements
NYC
Singapore
13,500 cabs
26,000 cabs
Shanghai, San Francisco, Vienna, ...
Page 3
Step 1: Analyze dataNY 170,000,000 trips / year
Pickups Dropoffs
Page 4
Urban taxi systems
Pickups Dropoffs 7 days in 20 sec
Page 5
Trips could be combined
Page 6
Previous attempts at improvement
• Ride sharing
• Smart hailing
Page 7
Can we come up with a new system?
• More efficient
• Less emissions
• Affordable alternative
Page 8
Step 2: A new dispatch algorithm
Combine 2 trips
Page 9
Step 2: A new dispatch algorithm
Combine k trips “Taxi Limousine”
Page 10
Manhattan street network
4000 intersections9000 street segments
Extracted fromOpenStreetMap
Match GPS-coords of pickup/dropoff points with street intersections
Page 11
Dynamic pickup and delivery problems
T1
T2
T3
T4
Like traveling salesman with time constraints
Small systems solvable with linear programming
Large systems not
Yang, Jaillet and Mahmassani, Transp Sci 38 (2004)Berbeglia, Cordeau and Laporte, Eur J Op Res 202 (2010)
Marin, An Op Res 143 (2006)
Page 12
Shareability networks
k = 2T1
T2
T3
T4
T2T1
T3
T4
Page 13
Shareability networks
k = 2T1
T2
T3
T4
T2T1
T3
T4
Solution: maximum matching
Generalizable to k>2but unfeasible for k>3
Chandra and Halldorsson, J Alg 39 (2001)
Page 14
Satisfaction criterion
Maximum time delay Δ
Δ = 30 sec Δ = 60 sec
more tolerance = denser network = more sharing opportunities
Krings et al, EPJ Data Sci 1 (2012)
Page 15
Oracle vs. Online
Oracle: omniscient, best possible
T1
T2
Online: realistic, constrained by time window δ
δ
Set δ = 1min
Page 16
Step 2: A new dispatch algorithm
• Send destination request (via app)
• Wait δ min
• Receive sharing options
• Trip may be prolonged up to Δ min
How it works:
Consequences:
• Less traffic = less pollution etc
• Split costs for customers
Page 17
Step 3: Simulation results: MOST trips can be combined!
Only δ = 1 min initial waiting time needed!
Page 18
Online tool for interactive exploration
http://hubcab.org
(in development)
Page 19
Zoom into the data
Pickups Dropoffs
Page 20
Michael Szell [email protected]
Benedikt GroßJoey LeeEric BaczukCarlo RattiAndi Weiß (47Nord)Stefan Landsbeck (47Nord)
Research Visualization & Explorer
hubcabTaxi-sharing in New York City: A network-based approach
Paolo SantiGiovanni RestaStanislav SobolevskyCarlo RattiSteven Strogatz (Cornell)