04 Introduction to Agent-Based
Modelling (ABM) (Part I)
By NTU Complexity Institute
What is ABM?
Acronyms
- ABM: Agent-Based Modelling
- ABS: Agent-Based Simulation
- ABMS: Agent-Based Modelling and Simulation
What is ABM?
What is an agent?
- Autonomous decision-making entity (Bonabeau, 2002)
- Usually implemented as an AI software object
- ‘Proactive’ refers to the ability to set goals
- ‘Reactive’ is to respond to changes in environment
- ‘Interactive’ refers to working with or against other agents to achieve goals
Why Do We Use ABM?
- Natural description of problem in terms of agents
- System of agents is complex
- Strategic interactions
- Heterogeneity Not easy to build SD models
- Emergent collective behaviour
- ABM is a bottom-up approach
- Exploration
- Better appreciate impact of poorly understood interactions
- Prediction
- Scenario analysis for public policy
How to Build an ABM?
Ingredients
- Agent types (attributes such as gender, age, preferences, …)
- Action types (behaviours such as buying, selling, holding, …)
- Conditions
Observations
- Agent types ✔
- Action types ✔
- Conditions?
Interviews and surveys
- Agent types ✔
- Action types ✔
- Conditions ✔
Level of Detail
Number of agent types
- Can we group agents?
Number of actions
- Can we group actions?
Number of conditions
- Are there combinations of conditions?
Thomas Schelling
(1921–2016)
2005 Nobel Prize winner in economics
Simple Example of ABM: Thomas Schelling’s Model
Schelling Model with two colours: Final State with Happiness Threshold 30%
Rule: If fewer than fraction p of neighbours are
of the same colour, move to a random vacant
site.
Segregation persists even when p becomes
vanishingly small!
Segregated neighbourhoods
04 Introduction to Agent-Based
Modelling (ABM) (Part II)
By NTU Complexity Institute
How to Validate the
ABM?
- Computer program must do what the model prescribes
- Verification problem
- For both exploratory and predictive ABM
- E.g., In a simulation of Schelling’s model, an unhappy agent at r = (x, y) is identified:
- According to the model, the agent will move to a random vacant site r’ = (x’, y’). r = (x, y) will now
be vacant
- The program creates a new agent at r’ = (x’, y’) but unhappy agent at r = (x, y) remains
- Implementation error must be corrected
How to Calibrate the
ABM?
For ABM to be predictive, it must calibrated by using:
- Aggregate (macroscopic) data
- E.g., For Schelling’s model, we can measure an aggregate segregation index from real
neighbourhoods, then adjust parameter p in model for simulations to produce same aggregate
segregation index
- Mesoscopic data
- E.g., For Schelling’s model, Fourier transform n(x, y) for real neighbourhoods to get N0(kx, ky).
Then adjust parameter p in model until simulations produce N(kx, ky) that best fits N0(kx, ky)
- Microscopic data
- E.g., For Schelling’s model, track the relocation histories of individual households, then adjust
parameter p in model until simulations produce best matching set of simulated relocation
histories
Microscopic data is necessary if p is different for different households (heterogeneity).
Examples of Calibration
Show the graph of index of dissimilarity for Schelling’s model
Show real-world index of dissimilarity as horizontal line to back out p
How to Validate ABM
against Additional Data?
Testing calibrated ABM against additional data:
- Macroscopic data
- Mesoscopic data
- Microscopic data
Depending on purpose of ABM, the model can be calibrated against microscopic/mesoscopic data,
but validated against macroscopic data
- Training data can be more detailed and more intrusive compared to operational data
Examples of Complex
ABM
Feature the MRT + bus model developed by IHPC
04 Introduction to Agent-Based
Modelling (ABM) (Part III)
By NTU Complexity Institute
ABM for Policy
Assessment
Risk assessment under different scenarios
- Test prevention, intervention and mitigation before actual events occur
- Synthetic futures for foresight exercises
Mechanistic understanding
- Change behaviours of agents to achieve desired outcomes and avert undesirable ones
Inform measurement of microscopic variables
- Identify the critical microscopic variables for given outcome
- Measure these in real life
Examples of ABM
Platforms
NetLogo: [Link]
Swarm: [Link]
Repast: [Link]
MASON: [Link]
Acknowledgements
- Slide 7: Photo of Thomas Schelling, extracted from Flickr: [Link] by New America:
[Link] under CC BY 2.0: [Link]
- Slide 8: Image generated with Python codes. Retrieved April 17, 2017, from [Link]
implementation-of-schelling-model-in-python/144.
While every effort have been made to undertake due diligence to identify the original source of materials used in the preceding video, this was not always possible for all materials.
The video is used solely for educational purposes, and is hosted and deployed in a secured learning management systems accessed by Coursera students with a user ID and password.
This disclaimer states that no copyright infringement was intended. Should any disagreement or dispute arise, including any claims to rightful ownership, kindly contact ldd_admin@[Link].