Agents for spatial modelling
Simulation
“ a la Magritte”
Agent Based Modelling
On the Earth
• Multiple interacting system types at different space and time scales
• Everything is connected to everything else!
Systems are generally
• Heterogeneous
• Spatially distributed
• Many systems contain interacting discrete objects
Dynamics are generally complex (not just complicated!)
•
•
•
•
•
Sensitive to initial/boundary conditions
Path-dependent/contingent/adaptive
Non-decomposable with multiple feedback loops
Far from equilibrium
Tipping points/Phase changes
Statistics are typically
• Non gaussian - often with fat-tails
• Non-stationary over time and/or space
Agent Based Modelling
Environmental Change typically displays complexity
•
•
•
•
•
•
Causes, processes and impacts of climate change
Ecosystems , their services, management and conservation
Environmental Hazards
Disease and pandemics
Urbanization and Land-use change
Economics, poverty, wealth distributions...
Agent Based Modelling
Environmental Change typically displays complexity
Explicitly Spatial Process-based understanding needed
needed
Environmental Models need to include the social process
•Dynamics of social, ecological and physical systems are coupled
•Decisions are not made on purely environmental but on economic, cultural and
political grounds also.
•Human and Ecological systems involve many interacting individuals in which global
structure emerges from and forms the framework for the small scale interactions
Large scale physical or economic simulations may seem remote, and may be hard to
translate into a form that relates to everyday experience.
•We need to make explicit the consequences of people's actions
•We need to give policy makers (and others) direct information about the
consequences of inaction.
Agent Based Modelling
Environmental Change typically displays complexity
Explicitly Spatial Process-based understanding needed
needed
Environmental Models need to include the social process
•Dynamics of social, ecological and physical systems are coupled
•Decisions are not made on purely environmental but on economic, cultural and
political grounds also.
•Human and Ecological systems involve many interacting individuals in which global
structure emerges from and forms the framework for the small scale interactions
Large scale physical or economic simulations may seem remote, and may be hard to
translate into a form that relates to everyday experience.
•We need to make explicit the consequences of people's actions
•We need to give policy makers (and others) direct information about the
consequences of inaction.
We typically lack explicit large-scale understanding
Agent Based Modelling
Environmental Change typically displays complexity
Explicitly Spatial Process-based understanding needed
We typically lack explicit large-scale understanding
needed
• Approach by simulation of every discrete object in the system
• Use sets of rules for behaviour that are intuitively
reasonable to model the human aspects of the
system
• Hope that large scale systematic behaviours emerge!
Agent Based Modelling
Agents are discrete actors capable of self-generated
autonomous activity that effect change in their own
state and/or that of their surroundings
Agent Based Modelling
Agents are discrete actors capable of self-generated
autonomous activity that effect change in their own
state and/or that of their surroundings
•A rock ?
• No
•A tree ?
• Yes
• reacts to light/water/CO2
• responds to attack with chemical signals
•An animal ?
• Yes
• adds possibility of free movement
• and reasoned behaviour – possibly even in humans...
Agent Based Modelling
Agents are discrete actors capable of self-generated
autonomous activity that effect change in their own
state and/or that of their surroundings
Agent-based models represent some aspects of real-world
agents using autonomous software components (objects).
•Originate in knowledge-based Artificial Intelligence
•Interlinked sets of rules determine agent behaviour
Agent Based Modelling
Agents are discrete actors capable of self-generated
autonomous activity that effect change in their own
state and/or that of their surroundings
Agent-based models represent some aspects of real-world
agents using autonomous software components (objects).
•Originate in knowledge-based Artificial Intelligence
•Interlinked sets of rules determine agent behaviour
Many types are possible in a single model
Trees, people, households, businesses, institutions, NGOs, governments,
cities, countries...
Many aspects possible, including qualitative information
emotions, goal directed behaviour, group formation, disease, crime...
Agent Based Modelling
Software agents are discrete entities with individually
modifiable properties and behaviours
•Agents are embedded in an environment
• Agents may be able to change the environment
• The environment may have its own dynamical processes that drive
change
• Environment may be continuous (e.g. topography) or discrete (rocks) or
some mixture of both
• Spatial agents additionally have spatial co-ordinates and may have the
power to change location within the environment
•An agent model typically has many agents of different types
– Often called Multi-Agent Systems (MAS)
– An agent may interact with and modify other agents, either directly or
indirectly through the environment
•In either case interaction can
– Imply complex feedback loops across time and space scales
– Lead to the emergence of structure not explicitly represented in any single agent
Agent Based Modelling
Software agents are discrete entities with individually
modifiable properties and behaviours
•Agents may be
“strong” with a rich cognitive structure, goal oriented plans, learning, norm
creation and internal representations of other entities
– able to reason about inputs and own internal state – can represent and model
their own surroundings, and learn new responses
“weak” or “reactive” with simple fixed reactions to other agents and the
environment, possibly in differing ways in similar circumstances depending on
history
Agent Models and the environment
Agent Based Modelling
•Advantages
– We can deal with heterogeneous, non equilibrium systems with
non-stationary time evolution
– Do “what-if” experiments in a way that may be impossible (or
unethical) in a real system
– Easier to communicate results to policy makers or to the wider
public
•Account for detailed population structure and behaviour
–
–
–
–
Age, sex, social class, health, wealth
Modify behaviour based on perceptions and memory
Bounded knowledge, erroneous beliefs, deception and deceit
Generate histories and examine path dependence
•Deal with “Space and place”
– semantic content
– movement, clustering, real geographies
Agent Based Modelling
Discrete Element Models
Avalanches and debris flows
Cliff-scree systems
Individual-Based Models
Forest simulation
Herds and flocking
Foraging
Predator-prey models
Agent-Based Models
Epidemics
Traffic simulation
Crowds and escape from disaster
Urban populations
Social-Ecological Systems
Land-use Change
Agent Based Modelling
Increasing
Complexity
Discrete Element Models
Avalanches and debris flows
Cliff-scree systems
Individual-Based Models
Forest simulation
Herds and flocking
Foraging
Predator-prey models
Agent-Based Models
Epidemics
Traffic simulation
Crowds and escape from disaster
Urban populations
Social-Ecological Systems
Land-use Change
Increasing
Numbers
Discrete Element Modelling
Cliff –scree systems
Discrete Element Modelling
Cliff-scree systems – distribution of avalanches
Early evolution – left-skewed, short tail, characteristic size
Later evolution – long tail, power law – self-organized criticality
Results largely independent of model parameters
Early period statistics –
a short tail and a modal value
Late period distributions - long tail,
no typical size, distribution is close to
a power law
Increasing
Complexity
Discrete Element Models
Avalanches and debris flows
Cliff-scree systems
Individual-Based Models
Forest simulation
Herds and flocking
Foraging
Predator-prey models
Agent-Based Models
Epidemics
Traffic simulation
Crowds and escape from disaster
Urban populations
Social-Ecological Systems
Land-use Change
Increasing
Numbers
Individual Based Modelling
Forest Model
•Individual-based model
representing each tree
•Allometric rules for tree
growth
•Competition primarily
through shading
•Different functional types
with varying shade tolerance
and growth parameters
•Examples
• SORTIE (Pacala et al 1996)
• TROLL(Chave 1999)
Individual Based Modelling
Forest Model
Low growth trees
in shade have a
high probability of
dying
Individual Based Modelling
Forest Model
Huge number of seedlings – but they don’t make it to maturity
Most of the wood is in large trees
Number by area distribution is exponential
Shade tolerant species crowd out the pioneers
Individual Based Modelling
Forest Model – a specific example
•Bore Khola Valley
•Nepal Middle Hills
•27.5º50’N 85º20’E
•20km N of Kathmandu
Individual Based Modelling
Forest Model – a specific example
•Bore Khola Valley
•Nepal Middle Hills
•27.5º50’N 85º20’E
•20km N of Kathmandu
Individual Based Modelling
Forest Model – a specific example
•4km square catchment
•Height data at 20m horizontal
resolution, 10m in vertical
Topography
•Overall relief approx. 1000m
4
km
N
Individual Based Modelling
Forest Model – a specific example
4
km
N
Year 0
•Tree density
accumulates over
time
Year 600
Year 1200
Individual Based Modelling
Forest Model – a specific example
4
km
N
Year 0
•Tree density
accumulates over
time
Year 600
Year 1200
•Now we send
people out into the
forest
Agent Based Modelling
Foraging
but memory gives a huge benefit in
gathering efficiency
with memory
random
Agent Based Modelling
Forest Model – a specific example
4
km
N
Year 0
•Tree density
accumulates over
time
Year 600
Year 1200
•Now we send
people out into the
forest
•Then people clear
the forest for
farming
Agent Based Modelling
Forest Model – a specific example
4
km
N
Year 0
•Tree density
accumulates over
time
Year 600
Year 1200
•Now we send
people out into the
forest
•Then people clear
the forest for
farming
Agent Based Modelling
Forest with people
•Fields Highlighted in
green, degraded forest
in yellow
•Farmers Exploit the
lower part of the
catchment first
•Trees are removed
much faster then they
can recover
Year 600
Year 660
Year 720
Agent Based Modelling
Forest with people
•Forested areas are
good at attenuating
water
•Soil compaction in
farmed areas increases
soil saturation
•As farming increases,
flash floods become
more likely
Year 600
Bithell and Brasington 2008
Year 660
Year 720
Agent Based Modelling
Increasing
Complexity
Discrete Element Models
Avalanches and debris flows
Cliff-scree systems
Individual-Based Models
Forest simulation
Herds and flocking
Foraging
Predator-prey models
Agent-Based Models
Epidemics
Traffic simulation
Crowds and escape from disaster
Urban populations
Social-Ecological Systems
Land-use Change
Increasing
Numbers
Agent Based Modelling
Agent Based Modelling
Canvey Island
In the great 1953
flood, sea defences
failed.
58 people died,
11000 evacuated.
Thames Estuary
Now 38000 people behind 4.66m high wall.
Only a single exit road
The illusion of safety provided by the sea wall
may have encouraged settlement.
Map data: Crown Copyright/ Database right 2013 – An Ordnance Survey/EDINA supplied service. Flood modelling by James Brown
Brown, J.D, Spencer,T. And Moeller,(2007) Water Resources Research 43.
Agent Based Modelling
Canvey Island
In the great 1953
flood, sea defences
failed
58 people died,
11000 evacuated
As the climate changes, extreme events are
predicted to become more likely.
Storm surges may again over- top or breach
the flood barrier
Simulated
breach
Map data: Crown Copyright/ Database right 2013 – An Ordnance Survey/EDINA supplied service. Flood modelling by James Brown
Brown, J.D, Spencer,T. And Moeller,(2007) Water Resources Research 43.
Agent Based Modelling
The island can flood
in a few hours
Evacuation may be
necessary
Traffic simulations
can help to
understand the
evacuation process
Map data: Crown Copyright/ Database right 2013 – An Ordnance Survey/EDINA supplied service. Flood modelling by James Brown
Brown, J.D, Spencer,T. And Moeller,(2007) Water Resources Research 43.
Agent Based Modelling
Policy options can
be tested to see
what might
improve evacuation
times.
. Ozioma Uzoegwu 2013 Mphil. Thesis
Agent Based Modelling and Disease
Disease models
Agents allow us to disentangle behaviour from other effects
•Propagation of disease is directly modelled as
transmission between infected and susceptible individuals
•Contact processes
– Can include effects of social networks
– Are constrained by the physical environment
– Can change with agent perceptions of symptoms
Agent Based Modelling and Disease
Agent Based Modelling
Conclusions
We can directly model processes in systems of discrete objects
Deal with situations where we lack analytic power
Emergent properties arise from collective interactions
Multiple coupled systems can be dealt with
Test policy options where not possible to experiment
Very visual – good for policy communication
Future
Larger scale, more complete, more complex systems
Social processes and networks in real-world situations
Model the “Anthropocene” –
current “Earth System Models” do not include people
 Grand Unified Models!
Agent Based Modelling
Agents programming systems and references
Modelling Environments
Netlogo
RePast (Swarm)
Mason
Gama
...many more
Netlogo is treated in detail in “Agent based and individual-based modelling: a
practical introduction” (2012) Grimm and Railsback (Princeton)
For a review of many others see “Design of Agent-based models” (2011) by
Tomas Salamon (Academic Series)
Starting references
•“Agent based models of Geographic Systems” 2012 Heppenstall et al eds. (springer)
•“Simulating Social Complexity: A Handbook” 2013 Edmonds and Meyer (eds) (springer)
•“An introduction to multi-agent systems” 2nd ed. 2009 Wooldridge (Wiley)
•“Growing artificial societies: social science from the bottom up” 1996 Epstein and Axtell
(brooking Institution Press) ... a classic!
Agent Based Modelling
Challenges
Model coupling
Cross-disciplinarity
Sharing and reproducing models/results
Joining complex dynamical models
Model ownership
Democratization of knowledge
Policy assessment
Risk and environmental change
Vizualization
System size
Spatial extent
Complex interacting dynamical systems
Problem framing
What should be modelled?
Who for?
What is relevant?
Data Gathering
How to understand behaviour
How to generalise case studies
How deal with large and small
scales
Scaling
System size
Parameter space exploration
Processes at different scales
Micro-macro links
Validation
Reflexivity
Causality
Data integrity
Handling uncertainty
Complexity
How intelligent do agents need to be?
How much complexity is “enough”?
What can be simulated?