The future of policy with agent-based models
Arnau Quera-Bofarull
Institute for Computational Cosmology, Durham, UK
How I got involved into ABM
- PhD student in computational Astrophysics
- March 2020: Royal society call for modelers (RAMP)
- JUNE project starts
- 20+ collaborators, input into
- Development continues:
long term plans with PHE
JUNE
- 56 million agents
- Resolution ~ English census
- Millions of venues
- Detailed customizable policies
Model fitting
We have
Model (tens of free parameters)
+
Data (quite messy)
It takes ~400 CPU hours to run one simulation....
How do we fit it?
Emulator
+
History Matching
Emulator + history matching
Train emulator
Run emulator
O(500k) times
Run full simulation O(100) times
Narrow parameter space search
Why we need the complexity
JUNE reproduces infection disparities among various demographic groups thanks to its granularity.
My contributions to JUNE
1) Main modeller, written ~60% of the code (25,000 lines)
2) Designed and implemented HPC parallelization structure.
3) Setup pipeline for model calibration against data
Interdisciplinary collaboration
1) Main modeller, written ~53% of the code (25,000 lines)
2) Designed and implemented HPC parallelization structure.
3) Setup pipeline for model calibration against data
The challenge of model initialization
Model calibration
Initial conditions
Latent (un-observable) variables.
Input parameters
JUNE latent variables
Observable
\theta_i
Simulator
latent variables include:
- asymptomatic carriers
- number of contacts
- compliance to policies
- current prevalence
Current emulation
Etalumis
Atılım Güneş Baydin,
Oxford
Calibration: A new approach
Etalumis
(Atılım Güneş Baydin, Oxford)
arXiv:1907.03382v2
Interface the simulator with a Probabilistic Programing Language
Etalumis
(Atılım Güneş Baydin, Oxford)
JUNE latent variables
Observable
Etalumis inference
We can reconstruct the microstate in JUNE.
This is crucial for policy making.
- Which schools can reopen?
- In which postcodes movement should be restricted?
- What are the most vulnerable hospitals?
- ...
Not limited to Epidemiology
JUNE is essentially a model of the English population and its dynamics
- Housing policy
- Public transport policy
- Origin of inequalities
- etc.
Challenging task, requires wide collaboration
Machine Learning
HPC
Specific domain knowledge
New programming languages
Bayesian statistics
Effective communication
Agent-based modelling
By arnauqb
