Simulation of Traffic Congestion as Complex Behaviour Society of Cartographers Annual Conference – 3 rd September 2012 Ed Manley Department of Civil, Environmental and Geomatic Engineering University College London
Dec 04, 2014
Simulation of Traffic Congestion as Complex Behaviour
Society of Cartographers Annual Conference – 3rd September 2012
Ed Manley Department of Civil, Environmental and Geomatic Engineering
University College London
Today’s Talk The Complexity of Road Congestion
• Behaviour and Complexity in
the City
• Agent-based Modelling of
Choice Behaviours
• Analysis of Taxi Driver Route
Selection Data
Urban Complexity A Product of Human Behaviour
• The function and nature of the city is defined by its the
choices of its citizens
• Choices influence how we interact
• This accumulation of behaviours lead to the patterns of
movement we see everyday
• Understanding and modelling these patterns requires a
fundamental understanding of human behaviour
Urban Complexity Road Congestion
• Road congestion is an excellent example of how human
behaviour influences urban dynamics
• People unilaterally pick their route and proceed towards
their target, they remain reactive to problems
• Competition for limited space at a given time results in
emergence of congestion
• Following shocks to the system, the influence of
individual responses is of greatest significance
Urban Complexity Understanding Individual Movement
• We examine the individual behaviours that contribute
towards the formation and spread of congestion
• How do drivers really choose a route?
• What areas of the city do they know best?
• How do they use information to aid them?
• What is the heterogeneity in behaviour across the
population?
• These behaviours are incorporated within an agent-
based model of the urban road system
Agent-based Modelling From Micro to Macro
• Agent-based Modelling allows us to link individual
behaviour with the macroscopic evolution of the system
• Individuals are represented distinctly, enabling
incorporation of population heterogeneity
• Individuals are autonomous and independent
• Interactions between agents may lead to emergence of
macroscopic phenomena
Case Study Investigating the Influence of Behaviour
• Aim to identify how different definitions of route selection
behaviour alter resulting road network patterns
• A range of individual route selection behaviours are
incorporated into agent-based model
Route Selection
Least Distance
Least Time
Least Angular
Least Turns
Spatial Knowledge
500m Area
1000m Area
Around OD Locations
Agent Behaviour Design
Driver agents independently choose route through city
Model Test Area Central London
Location: Central London
All road links
Road regulations and capacities integrated
30 minutes during AM peak
Agents: ~15000 driver agents
AM peak OD distribution
from TfL Trip Matrix
Model: Developed using Java +
Repast Simphony 1.2
© OpenStreetMap 2012
The Base Case
Base Case Path: Shortest Distance
Knowledge: Complete
1 0 0.5
mile
The Influence of Route Choice
Least Time Path: Least Time
Knowledge: Complete
Faster, main routes
Reduced on subsidiaries
Stronger influence in West
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
1 0 0.5
mile
Least Angular Path: Least Angular
Knowledge: Complete
Greater redistribution
Towards straighter sections
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
1 0 0.5
mile
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Least Turns Path: Least Turns (Distance Constrained)
Knowledge: Complete
Effect not as strong
Influenced by distance
But, highlights straighter sections
1 0 0.5
mile
The Influence of Spatial Knowledge
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Partial Knowledge Path: Shortest Distance
Knowledge: Reduced to 500m
Movement away from subsidiaries
Greater reliance on main routes
1 0 0.5
mile
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Partial Knowledge Path: Shortest Distance
Knowledge: Reduced to 1000m
Less deviation from base case
Reduction in use of subsidiaries
Due to greater all around knowledge
1 0 0.5
mile
Modelling Cities The Need for a Realistic Model of Behaviour
• Models demonstrate strong importance of establishing a
realistic representation of behaviour
• Small changes in behaviour definition lead to big
changes in city level patterns
• Establishing this model of behaviour represents an
important research goal
• In respect to route choice, we have been analysing route
trace data from minicab firm in London
Route Analysis Private Hire Cab Routes
• Dataset of 700k processed routes through London from
Addison Lee taxi company
• Not Black Cab drivers, but will have generally better
knowledge and may use navigation devices
• Analysis compared each route against a range of
optimal paths – here we will focus mainly on distance
• This work still in its early stages…
Taxi Driver Data Total Flows
Route Analysis Comparison to Alternatives – Averages
• For each whole route,
percentage of path
matched against range of
alternatives
• Average match taken for
each alternative
Choice Alternative Percentage
Matched
Least Distance 39.83
Least Time 38.21
Least Angular Deviation 27.37
Least Angular Deviation constrained by distance 33.06
Least Angular Deviation constrained by time 32.86
Least turns constrained by distance 42.48
Least right turns constrained by distance 39.48
Lowest descriptor term score constrained by distance 41.52
Lowest descriptor term score constrained by time 38.24
Lowest descriptor term score constrained by angle 28.58
Maximise number of lanes constraining distance 38.97
Maximise number of lanes constraining time 35.20
Maximise number of lanes constraining angle 25.47
Least turns constrained by time 39.50
Least right turns constrained by time 38.45
No strong stand out
artificial representation
of behaviour
Route Analysis Comparison to Alternatives – Good Matches
• Count of paths where
alternative matches over
75% of real journey
• Only journeys over 1km
in distance considered
Choice Alternative Percentage
Achieving 75%
Least Distance 13.1
Least Time 12.4
Least Angular Deviation 6.1
Least Angular Deviation constrained by distance 8.4
Least Angular Deviation constrained by time 8.8
Least turns constrained by distance 16.1
Least right turns constrained by distance 12.6
Lowest descriptor term score constrained by distance 15.9
Lowest descriptor term score constrained by time 13.2
Lowest descriptor term score constrained by angle 7.4
Maximise number of lanes constraining distance 12.8
Maximise number of lanes constraining time 10.7
Maximise number of lanes constraining angle 5.8
Least turns constrained by time 14.1
Least right turns constrained by time 12.7
Poor performance
by each measure of
prediction
WHY?
Route Analysis Spatial Distribution
• No complete routing algorithms provide an adequate
representation of reality
• This finding goes against assumptions within many
conventional models of traffic simulation
• So, which parts of these journeys are a good match
against optimal routes?
• We looked at deviations in route patterns across space,
by direction of travel, against optimal distance journeys
East to West London Journeys
Difference in flows between 7576 actual and
optimal distance routes
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Std. Dev. = 137.2
Mean = 4.1
Maximum = 1991
Minimum = -2365
1 0 0.5
mile
West to East London Journeys
Difference in flows between 9850 actual and
optimal distance routes
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Std. Dev. = 143.9
Mean = 4.5
Maximum = 1553
Minimum = -3018
1 0 0.5
mile
SE16 to W London Journeys
Difference in flows between 522 actual and
optimal distance routes
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Std. Dev. = 18.2
Mean = 1.3
Maximum = 130
Minimum = -176
1 0 0.5
mile
W to SE16 London Journeys
Difference in flows between 704 actual and
optimal distance routes
> 2.5 Std. Dev.
1.5 to 2.5 Std. Dev.
0.5 to 1.5 Std. Dev.
0.5 to -0.5 Std. Dev.
-0.5 to -1.5 Std. Dev.
-1.5 to -2.5 Std. Dev.
< -2.5 Std. Dev.
Std. Dev. = 27.4
Mean = 1.0
Maximum = 184
Minimum = -381
1 0 0.5
mile
Route Analysis Spatial Distribution
• Differences seem to indicate an attraction and
repulsion of certain parts of the road network
• Apparent preference for straight, longer sections,
possibly with greater salience or perception of travel time
• Route choice appears to not consist of a single route
selection, but a phase-based process of selection
• But does this mean distance plays no role at all? That
doesn’t appear to be quite the case…
Route Analysis Distance Minimisation
Route Analysis Choice Heterogeneity
• Indications are that route selection is a heuristic process,
probably involving minimisation of distance and route
complexity
• There is also a heterogeneity in decision-making –
Perhaps variation in knowledge? Location of decision?
• Analysing collections of paths between discrete locations
reveal that both of these factors may further contribute
E14 to Kings Cross Journeys
Flows of 521 routes between origin and
destination
1 0 0.5
mile
SE16 to W Journeys
Flows of 522 routes between
origin and destination
1 0 0.5
mile
W to SE16 Journeys
Flows of 704 routes between
origin and destination
1 0 0.5
mile
Route Analysis Decision Points
• Visualisations also allow us to identify locations of
significant splits in flow - decision points
• These areas of high activity are likely to be more salient
in an individual’s mind, on which choices made
• Decision points identified where inflow is split between
more than one outflow route (10% minimum)
• Could be used as foundation for decision making
process within model
E14 to Kings Cross Journeys
Decision Points origin and destination
Size indicates volume of traffic flow
through point
1 0 0.5
mile
Conclusions Summary of Research
• The definition of behaviour is clearly highly influential
in determining global patterns of movement
• Getting this representation right is key – requires full
examination of population heterogeneity
• Initial route analysis has highlighted some interesting
trends with relation to established assumptions
• Route choice appears to take place in phases
• Minimisation of distance and route complexity,
attraction to salient features appear important
Thank you
Ed Manley
Blog: http://UrbanMovements.posterous.com
Project: http://standard.cege.ucl.ac.uk
Twitter: @EdThink