SlideShare a Scribd company logo
1 of 51
Download to read offline
S I M U L AT I N G A N D O P T I M I S I N G A D E L I V E RY N E T W O R K O F
A U T O N O M O U S C O U R I E R A G E N T S
J A K E C R A C K N E L L – B E N G C O M P U T I N G – 2 0 1 5
S P E C I A L T H A N K S T O D R . K RY S I A B R O D A , P R O F. M A R E K S E R G O T A N D D R . T O N Y F I E L D
S O U R C E C O D E AVA I L A B L E AT G I T H U B . C O M / J A K E C R A C K N E L L
P D F O F T H E S I S AVA I L A B L E AT T I N Y. C C / C O U R I E R
M O T I VAT I O N
• Consumers increasingly demand same-day delivery
• Couriers are expensive
• Autonomous cars are on the horizon
• There may exist a gap in the market
P R O D U C T P R O P O S A L
• A distributed system of
unmanned, autonomous vehicles
• For businesses and consumers
• Jobs can ‘piggyback’ on others
• Potentially a lot cheaper
“I need a 50x50x30cm box picked up from SW7
2AZ and delivered to WC2R 2LS before 5pm”
+ 0 . 0 7 5 M 3
– 0 . 0 7 5 M 3
1 7 : 0 0 : 0 0
C I T Y S P R I N T: £ 1 3 . 9 5 + VAT
We have n cars with v m3 capacity
We are allowed to refuse some jobs
In some cases the sender or recipient will not be in
We want to minimise cost and jobs arriving late
T H E D E L I V E R A B L E S
• A multi-agent courier network simulation in
large cities with traffic
• New algorithms to efficiently fulfil jobs
(routing, planning and allocation)
• An evaluation of the sensitivity of parameters
• A feasible pricing model
D E M O
G E O G R A P H Y
• OpenStreetMap
• Nodes and Ways
• Hash-based adjacency list
• Spatially indexed array
• Mercator map projection
M A P P R U N I N G
• OSM has mapping errors and
omissions
• Agents mustn’t get trapped
• Road network should be
strongly connected
• I adapted Kosarju’s algorithm
Inescapable
Inaccessible
R O U T E F I N D I N G
• A* Algorithm
• One can minimise:
• Distance
• Time (with or without traffic)
• Fuel (with or without traffic)
• Routes are cached
• Bounded Relaxation
d is the straight-line distance from a node to the destination
max_speed is the speed limit over all roads
86 km/h is the best speed for maximising fuel economy
Bounded relaxation speeds up a search by increasing the weight of the
admissible heuristic function.
f(n) = g(n) + (1 + ε)•h(n)
A solution is guaranteed to be no worse than (1+ε) times longer than the
optimal.
ε=0
ε=3
Trade-off:
Increasing ε substantially improves execution time
Increasing ε leads to longer, less optimal routes
T R A F F I C F L O W
• HERE Maps provides live traffic
flow data via API
• For one week, samples were taken
every 5 minutes
• Difficult to integrate with OSM
• Greater London, Isle of Wight and
Las Vegas Valley
R O A D D E L AY S
• Obstacles that make a vehicle stop
for a period of time.
• They activate at different times. Some
oscillate between time periods,
others activate using a probabilistic
model.
• Some are mapped in OSM, some are
‘unexpected delays’.
D I S PAT C H E R
• A centralised component
• Generates B2B, B2C, C2C and
C2C jobs
• OSM mapped businesses
• Generates size and deadline
O P T I M I S AT I O N
A L L O C AT I O N
P L A N N I N G
R O U T E F I N D I N G
T H E T R AV E L L I N G S A L E S M A N P R O B L E M
T H E T R AV E L L I N G S A L E S M A N P R O B L E M
265252859812191058636308480000000 possible solutions
T H E C O U R I E R P L A N N I N G P R O B L E M
• The travelling salesman problem with extra constraints
• Precedence – pick-ups must be visited before drop-offs
• Capacity – vehicles cannot overfill
• Deadlines – drop-offs have delivery deadlines
• Fuel – vehicle cannot run out of fuel mid-route
• Some early waypoints may be ‘locked’ in place
• Arc costs can change throughout the day
• n waypoints → n! possible solutions
• The goal is to minimise:
• The number of deliveries that arrive late
• The total driving cost
• Solved using
• A greedy algorithm (faster, less optimal)
• A genetic algorithm (slower, more optimal)
T H E C O U R I E R P L A N N I N G P R O B L E M
G E N E T I C A L G O R I T H M
1 0 1 2 A B C D E F 1 0 1 2 A B C D E F
1 0 1 4 B A C D E F
1 0 1 5 A B C E D F
1 0 1 9 A B C D F E
1 0 1 9 A C B D E F
1 7 B C A D E F
1 0 1 2 A B C D E F
1 0 1 4 B A C D E F
1 0 1 5 A B C E D F
1 0 1 5 B A D C E F
1 0 1 9 A B C D F E
1 0 1 9 A C B D E F
1 0 2 1 B A C E D F
Return top
solution
from poolx500
A nearest-neighbour route, which might not be on-time
Score
A D J A C E N T PA I R W I S E M U TAT I O N S
• Adjacent waypoints wi → wj can be swapped in these cases:
• They are both drop-offs or both pick-ups
• wi is a pick-up, wj is a drop-off and they are not the same job!
• wi is a drop-off, wj is a pick-up and there would be space for wj at i.
T I M E 1 2 3 4 5 6
T Y P E P I C K P I C K D R O P D R O P P I C K D R O P
J O B A B B A C C
✘✔ ✔✘ Depends
T H E F I T N E S S F U N C T I O N
• Cost of the route, measured in kilometres, hours or litres of fuel
• A large penalty for each late delivery
• An approximate cost function is used due to time constraints
1000*LateDeliveries + ∑ cost(r)
r ∈ Routes
P E R F O R M A N C E
1 1.5 2
m Route Score
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.5 1 1.5 2
GENETICRouteScore
Optimum Route Score
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.5 1 1.5 2 2.5 3 3.5 4
GENETICRouteScore
GREEDY Route Score
Greedy
Better
Genetic
Better
• 2500 problems consisting of 4 jobs (8
waypoints) were generated
• Compared to the actual optimum:
• Greedy Algorithm: +43.7% longer
• Genetic Algorithm: +34.1% longer
• Is this good enough?
P E R F O R M A N C E
Enumerate all solutions and you’ll typically find it in the top 0.5%.
You want a solution that minimises the number of late deliveries?
A L L O C AT I O N
• Jobs are statically allocated to agents using the contract net protocol
• Agents bid extra driving cost. If no agent bids, the job is refused.
• A distributed problem solver. Local optimisation performed in parallel.
The black agent is awarded the job, customer charged for 15 minutes
C N P VA R I A N T S
• A number of variants were written, which vary in flexibility
C N P 5
• Jobs are tentatively allocated to
agents
• If circumstances change for the
worse and 1+ jobs may arrive late:
• Enumerate all unpicked jobs;
order them by their urgency
• In sequence attempt to auction
them off to other agents
R E F U E L L I N G
• Agents burn fuel at different
rates, at different speeds
• If their fuel tank drops to 5%,
they postpone jobs and refuel
• Could this be done during off-
peak times?
I D L E S T R AT E G I E S
• What should an agent do
when it has no jobs:
• Strategic placement
• Premature refuelling
• Cost overhead of ~6%
• Typically at night
S E
F
F
F
FF
F
F
E VA L U AT I O N
D I S PAT C H R AT E
V E H I C L E T Y P E
V E H I C L E T Y P E
MilesPerGallon = k × (−0.0119x2
+ 1.2754x)
V E H I C L E T Y P E
0
0
10 km
6.2 mi
T H E O P T I C A L N E T W O R K S I Z E
A S S U M I N G ~ 2 0 J O B S P E R H O U R AT P E A K T I M E
2 0
4
5
9
4 7
C O S T C O M PA R I S O N F O R G R E AT E R L O N D O N
• Caution: Only fuel costs monitored
• Caution: Very dependent on parameters
• Break even pricing model: £2 + £2.37/hour
• Average customer paid £3.81
• £4 + £5/hour led to £5721.13/week profit
• CitySprint and Shutl prices are £10–100s
£
F U T U R E W O R K
• New vehicles, dispatchers and idle strategies
• Progressive planning and route finding
• Handoff/exchange at depots
• Map-specific strategies
S TA G E A * ε G E N S C O S T
I N I T 5
G R E E D Y
O N LY
3 . 2
1 2 5 0 0 2 . 7
2 1 1 0 0 0 2 . 5
3 0 . 5 1 5 0 0 2 . 4
4 0 ∞ 2 . 3
S U M M A RY
• Rich simulation with real maps and traffic
• Flexible route finding
• Greedy and genetic planning algorithms
• Routing and allocation strategies, namely CNP
• Idle strategies
• Extensive evaluation suggests very low costs
S O U R C E C O D E AVA I L A B L E AT G I T H U B . C O M / J A K E C R A C K N E L L
P D F O F T H E S I S AVA I L A B L E AT T I N Y. C C / C O U R I E R
S I M U L AT I N G A N D O P T I M I S I N G A D E L I V E RY N E T W O R K O F
A U T O N O M O U S C O U R I E R A G E N T S
J A K E C R A C K N E L L – B E N G C O M P U T I N G – 2 0 1 5
S P E C I A L T H A N K S T O D R . K RY S I A B R O D A , P R O F. M A R E K S E R G O T A N D D R . T O N Y F I E L D
Appendix
AACourier-30m-Industry
AACourier-30m-Industry
AACourier-30m-Industry

More Related Content

Similar to AACourier-30m-Industry

4Developers 2015: Measure to fail - Tomasz Kowalczewski
4Developers 2015: Measure to fail - Tomasz Kowalczewski4Developers 2015: Measure to fail - Tomasz Kowalczewski
4Developers 2015: Measure to fail - Tomasz KowalczewskiPROIDEA
 
Strangler Pattern in practice @PHPers Day 2019
Strangler Pattern in practice @PHPers Day 2019Strangler Pattern in practice @PHPers Day 2019
Strangler Pattern in practice @PHPers Day 2019Michał Kurzeja
 
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...ivaderivader
 
Dynamic demand
Dynamic demandDynamic demand
Dynamic demandJumpingJaq
 
Tesla y Dassault Systèmes unidos de la mano
Tesla y Dassault Systèmes unidos de la manoTesla y Dassault Systèmes unidos de la mano
Tesla y Dassault Systèmes unidos de la manoJaime Requeno
 
Robotics:The computational motion planning:: Sampling based algorithms
Robotics:The computational motion planning:: Sampling based algorithmsRobotics:The computational motion planning:: Sampling based algorithms
Robotics:The computational motion planning:: Sampling based algorithmsPRAVEENTALARI4
 
Using satellite imagery to track economic change
Using satellite imagery to track economic changeUsing satellite imagery to track economic change
Using satellite imagery to track economic changeRishabh Srivastava
 
The Great Policy Smog and how to navigate it
The Great Policy Smog and how to navigate itThe Great Policy Smog and how to navigate it
The Great Policy Smog and how to navigate itEMEX
 
PAC 2020 Santorin - Joerek Van Gaalen
PAC 2020 Santorin - Joerek Van GaalenPAC 2020 Santorin - Joerek Van Gaalen
PAC 2020 Santorin - Joerek Van GaalenNeotys
 
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareShuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareAsadullah Malik
 
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareShuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareAsadullah Malik
 
Velocity NY 2014 - The Natives are Getting Restless
Velocity NY 2014 - The Natives are Getting RestlessVelocity NY 2014 - The Natives are Getting Restless
Velocity NY 2014 - The Natives are Getting RestlessCliff Crocker
 
Driver Productivity: Where the Rubber Meets the Road
Driver Productivity: Where the Rubber Meets the RoadDriver Productivity: Where the Rubber Meets the Road
Driver Productivity: Where the Rubber Meets the RoadBarcoding, Inc.
 
Stop405tolls.org presentation to WA State House Transportation Committee
Stop405tolls.org presentation to WA State House Transportation CommitteeStop405tolls.org presentation to WA State House Transportation Committee
Stop405tolls.org presentation to WA State House Transportation CommitteeDavid Hablewitz
 
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...Денис Киркач
 

Similar to AACourier-30m-Industry (20)

publieke_nomovie
publieke_nomoviepublieke_nomovie
publieke_nomovie
 
4Developers 2015: Measure to fail - Tomasz Kowalczewski
4Developers 2015: Measure to fail - Tomasz Kowalczewski4Developers 2015: Measure to fail - Tomasz Kowalczewski
4Developers 2015: Measure to fail - Tomasz Kowalczewski
 
Measure to fail
Measure to failMeasure to fail
Measure to fail
 
Strangler Pattern in practice @PHPers Day 2019
Strangler Pattern in practice @PHPers Day 2019Strangler Pattern in practice @PHPers Day 2019
Strangler Pattern in practice @PHPers Day 2019
 
Technology Overview
Technology OverviewTechnology Overview
Technology Overview
 
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...
Augmenting Decisions of Taxi Drivers through Reinforcement Learning for Impro...
 
Dynamic demand
Dynamic demandDynamic demand
Dynamic demand
 
Tesla sf tm
Tesla sf tmTesla sf tm
Tesla sf tm
 
Tesla y Dassault Systèmes unidos de la mano
Tesla y Dassault Systèmes unidos de la manoTesla y Dassault Systèmes unidos de la mano
Tesla y Dassault Systèmes unidos de la mano
 
Everybody Lies
Everybody LiesEverybody Lies
Everybody Lies
 
Robotics:The computational motion planning:: Sampling based algorithms
Robotics:The computational motion planning:: Sampling based algorithmsRobotics:The computational motion planning:: Sampling based algorithms
Robotics:The computational motion planning:: Sampling based algorithms
 
Using satellite imagery to track economic change
Using satellite imagery to track economic changeUsing satellite imagery to track economic change
Using satellite imagery to track economic change
 
The Great Policy Smog and how to navigate it
The Great Policy Smog and how to navigate itThe Great Policy Smog and how to navigate it
The Great Policy Smog and how to navigate it
 
PAC 2020 Santorin - Joerek Van Gaalen
PAC 2020 Santorin - Joerek Van GaalenPAC 2020 Santorin - Joerek Van Gaalen
PAC 2020 Santorin - Joerek Van Gaalen
 
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareShuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
 
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS softwareShuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
Shuttle Route Optimization for the Sector # H-12, Islamabad using GIS software
 
Velocity NY 2014 - The Natives are Getting Restless
Velocity NY 2014 - The Natives are Getting RestlessVelocity NY 2014 - The Natives are Getting Restless
Velocity NY 2014 - The Natives are Getting Restless
 
Driver Productivity: Where the Rubber Meets the Road
Driver Productivity: Where the Rubber Meets the RoadDriver Productivity: Where the Rubber Meets the Road
Driver Productivity: Where the Rubber Meets the Road
 
Stop405tolls.org presentation to WA State House Transportation Committee
Stop405tolls.org presentation to WA State House Transportation CommitteeStop405tolls.org presentation to WA State House Transportation Committee
Stop405tolls.org presentation to WA State House Transportation Committee
 
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...
tew (10.10.18) - Transmission tariff calculation for ukraine, using the capac...
 

AACourier-30m-Industry

  • 1. S I M U L AT I N G A N D O P T I M I S I N G A D E L I V E RY N E T W O R K O F A U T O N O M O U S C O U R I E R A G E N T S J A K E C R A C K N E L L – B E N G C O M P U T I N G – 2 0 1 5 S P E C I A L T H A N K S T O D R . K RY S I A B R O D A , P R O F. M A R E K S E R G O T A N D D R . T O N Y F I E L D S O U R C E C O D E AVA I L A B L E AT G I T H U B . C O M / J A K E C R A C K N E L L P D F O F T H E S I S AVA I L A B L E AT T I N Y. C C / C O U R I E R
  • 2. M O T I VAT I O N • Consumers increasingly demand same-day delivery • Couriers are expensive • Autonomous cars are on the horizon • There may exist a gap in the market
  • 3. P R O D U C T P R O P O S A L • A distributed system of unmanned, autonomous vehicles • For businesses and consumers • Jobs can ‘piggyback’ on others • Potentially a lot cheaper
  • 4. “I need a 50x50x30cm box picked up from SW7 2AZ and delivered to WC2R 2LS before 5pm” + 0 . 0 7 5 M 3 – 0 . 0 7 5 M 3 1 7 : 0 0 : 0 0 C I T Y S P R I N T: £ 1 3 . 9 5 + VAT
  • 5.
  • 6.
  • 7. We have n cars with v m3 capacity We are allowed to refuse some jobs In some cases the sender or recipient will not be in We want to minimise cost and jobs arriving late
  • 8. T H E D E L I V E R A B L E S • A multi-agent courier network simulation in large cities with traffic • New algorithms to efficiently fulfil jobs (routing, planning and allocation) • An evaluation of the sensitivity of parameters • A feasible pricing model
  • 9. D E M O
  • 10. G E O G R A P H Y • OpenStreetMap • Nodes and Ways • Hash-based adjacency list • Spatially indexed array • Mercator map projection
  • 11. M A P P R U N I N G • OSM has mapping errors and omissions • Agents mustn’t get trapped • Road network should be strongly connected • I adapted Kosarju’s algorithm Inescapable Inaccessible
  • 12. R O U T E F I N D I N G • A* Algorithm • One can minimise: • Distance • Time (with or without traffic) • Fuel (with or without traffic) • Routes are cached • Bounded Relaxation
  • 13. d is the straight-line distance from a node to the destination max_speed is the speed limit over all roads 86 km/h is the best speed for maximising fuel economy
  • 14. Bounded relaxation speeds up a search by increasing the weight of the admissible heuristic function. f(n) = g(n) + (1 + ε)•h(n) A solution is guaranteed to be no worse than (1+ε) times longer than the optimal.
  • 15. ε=0
  • 16. ε=3
  • 17. Trade-off: Increasing ε substantially improves execution time Increasing ε leads to longer, less optimal routes
  • 18. T R A F F I C F L O W • HERE Maps provides live traffic flow data via API • For one week, samples were taken every 5 minutes • Difficult to integrate with OSM • Greater London, Isle of Wight and Las Vegas Valley
  • 19.
  • 20. R O A D D E L AY S • Obstacles that make a vehicle stop for a period of time. • They activate at different times. Some oscillate between time periods, others activate using a probabilistic model. • Some are mapped in OSM, some are ‘unexpected delays’.
  • 21. D I S PAT C H E R • A centralised component • Generates B2B, B2C, C2C and C2C jobs • OSM mapped businesses • Generates size and deadline
  • 22. O P T I M I S AT I O N A L L O C AT I O N P L A N N I N G R O U T E F I N D I N G
  • 23. T H E T R AV E L L I N G S A L E S M A N P R O B L E M
  • 24. T H E T R AV E L L I N G S A L E S M A N P R O B L E M 265252859812191058636308480000000 possible solutions
  • 25. T H E C O U R I E R P L A N N I N G P R O B L E M • The travelling salesman problem with extra constraints • Precedence – pick-ups must be visited before drop-offs • Capacity – vehicles cannot overfill • Deadlines – drop-offs have delivery deadlines • Fuel – vehicle cannot run out of fuel mid-route • Some early waypoints may be ‘locked’ in place • Arc costs can change throughout the day • n waypoints → n! possible solutions
  • 26. • The goal is to minimise: • The number of deliveries that arrive late • The total driving cost • Solved using • A greedy algorithm (faster, less optimal) • A genetic algorithm (slower, more optimal) T H E C O U R I E R P L A N N I N G P R O B L E M
  • 27. G E N E T I C A L G O R I T H M 1 0 1 2 A B C D E F 1 0 1 2 A B C D E F 1 0 1 4 B A C D E F 1 0 1 5 A B C E D F 1 0 1 9 A B C D F E 1 0 1 9 A C B D E F 1 7 B C A D E F 1 0 1 2 A B C D E F 1 0 1 4 B A C D E F 1 0 1 5 A B C E D F 1 0 1 5 B A D C E F 1 0 1 9 A B C D F E 1 0 1 9 A C B D E F 1 0 2 1 B A C E D F Return top solution from poolx500 A nearest-neighbour route, which might not be on-time Score
  • 28. A D J A C E N T PA I R W I S E M U TAT I O N S • Adjacent waypoints wi → wj can be swapped in these cases: • They are both drop-offs or both pick-ups • wi is a pick-up, wj is a drop-off and they are not the same job! • wi is a drop-off, wj is a pick-up and there would be space for wj at i. T I M E 1 2 3 4 5 6 T Y P E P I C K P I C K D R O P D R O P P I C K D R O P J O B A B B A C C ✘✔ ✔✘ Depends
  • 29. T H E F I T N E S S F U N C T I O N • Cost of the route, measured in kilometres, hours or litres of fuel • A large penalty for each late delivery • An approximate cost function is used due to time constraints 1000*LateDeliveries + ∑ cost(r) r ∈ Routes
  • 30. P E R F O R M A N C E 1 1.5 2 m Route Score 0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 GENETICRouteScore Optimum Route Score 0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4 GENETICRouteScore GREEDY Route Score Greedy Better Genetic Better • 2500 problems consisting of 4 jobs (8 waypoints) were generated • Compared to the actual optimum: • Greedy Algorithm: +43.7% longer • Genetic Algorithm: +34.1% longer • Is this good enough?
  • 31. P E R F O R M A N C E Enumerate all solutions and you’ll typically find it in the top 0.5%. You want a solution that minimises the number of late deliveries?
  • 32. A L L O C AT I O N • Jobs are statically allocated to agents using the contract net protocol • Agents bid extra driving cost. If no agent bids, the job is refused. • A distributed problem solver. Local optimisation performed in parallel.
  • 33. The black agent is awarded the job, customer charged for 15 minutes
  • 34. C N P VA R I A N T S • A number of variants were written, which vary in flexibility
  • 35. C N P 5 • Jobs are tentatively allocated to agents • If circumstances change for the worse and 1+ jobs may arrive late: • Enumerate all unpicked jobs; order them by their urgency • In sequence attempt to auction them off to other agents
  • 36. R E F U E L L I N G • Agents burn fuel at different rates, at different speeds • If their fuel tank drops to 5%, they postpone jobs and refuel • Could this be done during off- peak times?
  • 37. I D L E S T R AT E G I E S • What should an agent do when it has no jobs: • Strategic placement • Premature refuelling • Cost overhead of ~6% • Typically at night S E F F F FF F F
  • 38. E VA L U AT I O N
  • 39. D I S PAT C H R AT E
  • 40. V E H I C L E T Y P E
  • 41. V E H I C L E T Y P E MilesPerGallon = k × (−0.0119x2 + 1.2754x)
  • 42. V E H I C L E T Y P E
  • 43. 0 0 10 km 6.2 mi T H E O P T I C A L N E T W O R K S I Z E A S S U M I N G ~ 2 0 J O B S P E R H O U R AT P E A K T I M E 2 0 4 5 9 4 7
  • 44. C O S T C O M PA R I S O N F O R G R E AT E R L O N D O N • Caution: Only fuel costs monitored • Caution: Very dependent on parameters • Break even pricing model: £2 + £2.37/hour • Average customer paid £3.81 • £4 + £5/hour led to £5721.13/week profit • CitySprint and Shutl prices are £10–100s £
  • 45. F U T U R E W O R K • New vehicles, dispatchers and idle strategies • Progressive planning and route finding • Handoff/exchange at depots • Map-specific strategies S TA G E A * ε G E N S C O S T I N I T 5 G R E E D Y O N LY 3 . 2 1 2 5 0 0 2 . 7 2 1 1 0 0 0 2 . 5 3 0 . 5 1 5 0 0 2 . 4 4 0 ∞ 2 . 3
  • 46. S U M M A RY • Rich simulation with real maps and traffic • Flexible route finding • Greedy and genetic planning algorithms • Routing and allocation strategies, namely CNP • Idle strategies • Extensive evaluation suggests very low costs
  • 47. S O U R C E C O D E AVA I L A B L E AT G I T H U B . C O M / J A K E C R A C K N E L L P D F O F T H E S I S AVA I L A B L E AT T I N Y. C C / C O U R I E R S I M U L AT I N G A N D O P T I M I S I N G A D E L I V E RY N E T W O R K O F A U T O N O M O U S C O U R I E R A G E N T S J A K E C R A C K N E L L – B E N G C O M P U T I N G – 2 0 1 5 S P E C I A L T H A N K S T O D R . K RY S I A B R O D A , P R O F. M A R E K S E R G O T A N D D R . T O N Y F I E L D