Presentation Outline

1. Digital twins (of urban systems)

2. The role of agent-based modelling

3. Challenges for Urban Digital Twins

4. Progress towards Urban DTs

Data, model synthesis, emulation, real-time modelling

5. Are they worth it?

What are Digital Twins?

A synthesis of computer models, sensor networks, visualisations, etc., that mirror a real-world system, product or process

"Precise, virtual copies of machines or systems"

Examples:

Machines / products;
Manufacturing (factories);
Health / hospitals;
Smart Cities (e.g. Singapore, Victoria, Bradford);
... many others ...

Urban Digital Twins

Recent significant interest from government (and industry / academia)

National Digital Twin programme

e.g. Gemini Principles

Connected Urban Twins (Hamburg, Leipzig, Munich)

Singapore

Bradford

Glasgow (University campus)

But lots of 3D models. Few twins

Urban Digital Twins

Potential (?)

Improved (public) transport & mobility

Rapid, evidence-based responses to evolving urban state (emergencies, organised events, etc.)

More effective use of front-line services

Agile, responsive policy making

Why ABM?

Emergence

"The whole is greater than the sum of its parts." (Aristotle?)

Simple rules → complex outcomes

E.g. who plans the air-conditioning in termite mounds?

Hard to anticipate, and cannot be deduced from analysis of an individual

ABM uses simulation to (try to) understand how macro-level patterns emerge from micro-level behaviours

Why ABM?

Individual-level Representations of Theory

Implement individual behavioural rules into agents

Representing Space

Rich micro-level environments

Natural Description of a System

Describe the entities directly, rather than using aggregate equations

Why ABM?

History of the Model Evolution

Rather than returning a single result, the model evolves

The evolution itself can be interesting

Analyse why certain events occurred

Agent-Based Modelling - Difficulties

Tendency towards minimal behavioural complexity

Stochasticity

Computationally expensive (not amenable to optimisation)

Complicated agent decisions, lots of decisions, multiple model runs

Modelling "soft" human factors

Need detailed, high-resolution, individual-level data

Individual-level data

The Role of Agent-Based Modelling in Urban DTs?

ABMs are ideal in some circumstances

Interactions are important

Individual heterogeneity

Lots of individuals (but not too many...)

But the wrong tool in other circumstances

If a simpler model works - use that!

(Spatial interaction models, (GW) regression, dynamic microsimulation, etc...)

Most useful role in a DT as part of a suite of models?

Challenges - Uncertainty

Many sources of uncertainty (Ghahramani, 2015; Edeling, et al., 2021):

Measurement noise / observation uncertainty

Parameter uncertainty

Related problems with identifiability / equifinality

Also ensemble variance (model runs give different outcomes)

Model structure uncertainty

Model discrepancy and over-fitting

Scenario uncertainty (for predictions)

Need to understand and be honest about uncertainties

Uncertainty Quantification can help (but relevant to DT models?)

Challenges - Data

Even in the Age of Data, there are huge unknowns

Models (DTs) can be extremely detailed

Individual people / objects, interactions, etc.

But typically only coarse data are available

How to calibrate and validate all model components?

c.f. Pattern Oriented Modelling (Grimm at al., 2005)

And in real time ?!?

Challenges

Computation and Model Synthesis

(Very!) computationally expensive models

Innovations in (e.g.) meteorology not necessarily applicable to urban DTs

Also technical challenges coupling models

Individual models are fragile, when combined ... ⚠

DTs are a "synthesis" of computer models

'Big' Data and Urban DTs

Pedestrian Mobility

Growth in availability of data for quantifying the 'ambient population' (Whipp et al., 2021)

Census

travel surveys

mobile phone activity

smart phone apps

social media

pedestrian counters (WiFi, CCTV)

Pedestrian Mobility

Melbourne pedestrian counters

40+ sensors collecting over 8+ years

ML (random forest) to estimate pedestrian volumes based on:

Local street configuration (betweenness)

Street furniture (benches, bins, lights...)

Time (day, week, season, etc.) (1-hour resolution)

Weather

...

Pedestrian Mobility

Errors (potentially) interesting

Suggest where footfall is more/less predictable

Model uses

Prospective: Predict current and future (short-term) footfall (out of sample prediction)

Retrospective: Quantify success of organised events

Theoretical: Build evidence base for drivers of population mobility

Part of an urban digital twin?

Model synthesis

DyME: Dynamic Model for Epidemics

COVID transmission model with components including:

dynamic spatial microsimulation, spatial interaction model, data linkage (PSM), ...

Represents all individuals in a study area with activities: home, shopping, working, schooling

Daily timestep

DyME Validation Drawback: Data

Incredible detailed model!

BUT only data available for validation: COVID cases and hospital deaths

Only quantify a tiny part of the transmission dynamics

Huge uncertainties

Building the model was the easy part ...

Progress: Uncertainty Quantification

Many sources of uncertainty. Leverage uncertainty quantification?

History Matching (HM)

Discards parameter regions found to be implausible, leaving a smaller non-implausible region

Approximate Bayesian Computation (ABC)

ABC estimates a posterior distribution that quantifies the probability of specific parameter values given the observed data

HM + ABC

Sugarscape: HM

Very well known ABM (agents move around and consume 'sugar')

Measure the implausibility of parameters metabolism and vision

10 waves performed – use plausible space from wave 1 as input to wave 2 and so forth

Sugarscape: ABC

Parameters with the highest probability of matching the observation (i.e. sustaining a population of 66 agents) are where {metabolism, vision} are {4, 7}.

HM and ABC advantages

Provides a region of parameters and their probabilities, after accounting for uncertainties in the model and data

Rule out implausible models and help to explain observed trends (e.g. post-Brexit cattle farm scenarios (McCulloch et al, 2022))

Overall: more information than point-estimation calibration methods (although still more computationally expensive)

Dynamic Calibration for DyME

Recall: Very complicated COVID microsimulation, but large uncertainties and limited data for validation

Use ABC and Bayesian updating ('dynamic re-calibration')

Reduce uncertainty and produce more accurate future predictions

Parameter posteriors might reveal information about the model / system

Aside: Probabilistic ABM

Aside: Probabilistic ABM

Computational Efficiency

Uncertainty quantification, etc., requires many, many model runs

Difficult with computationally-expensive models, like ABMs

DyME (COVID microsimulation)

A single model run (800,000 individuals, 90 days) took 2 hours

ABC etc. would be impossible at that speed (need 1000s of runs)

Big computers can help

But maybe if I were better at programming ...

Model Emulation

But what if we can't speed up the model?

Attempt to use emulators (aka 'surrogate models')

Simpler versions of the model that run quickly

Crowd modelling case study: use an ABM to generate data to train an emulator:

Regression emulator (Random Forest)

Time-series emulator (Long Short-Term Memory neural network)

Emulator then makes real-time predictions quickly and can predict directly from aggregate data

Model Emulation

Scenario: pedestrian movement along a long corridor with 10 sensors

Try to predict sensor counts and total delay

Data assimilation v.s. calibration

Challenges for using DA with ABMs

Model size

10,000 agents * 5 variables = 50,000 distinct parameters

Agent behaviour

Agent's have goals, needs, etc., so can't be arbitrarily adjusted

Assumptions and parameter types

Maths typically developed for continuous parameters and assume normal distributions

... but, at least, many of these problems are shared by climate models

Real Time City Crowd Modelling

Simulating a city in real-time is too hard!! (for now)

For now lets start a crowd

What methods can we use to incorporate data?

How much data do we need?

Track every individual?

Track some individuals?

Just aggregate counts (e.g. number of people passing a footfall camera)

Data assimilation with a Particle Filter

Particle Filter & Crowd Simulation

Crowd Simulation with a Particle Filter

Preliminary Particle Filter Results

Box Environment: More particles = lower error

Difficulties (I)

Exponential increase in complexity

Preliminary Particle Filter Results

Grand Central Station: Filtering makes it worse!

Entrance gate is known; speed and exit gate are unknown

Particle Filter & Categorical Parameters

(video)

Categorical-Noise PF Step

Overview

Categorical-Noise PF Step

Results

Other DA methods (i)

Unscented Kalman Filter (UKF)

Similar to the (very popular) Ensemble Kalman filter

Should be more efficient

But assumes Gaussian distributions

A few sigma points are chosen to represent the model state

Then some complicated maths happens ...

Unscented Kalman Filter (UKF)

Observe a proportion of the agents

Unscented Kalman Filter (UKF)

Observed v.s. Unobserved agents

Other DA methods (ii)

A new approach: Monte-Carlo Markov Chain sampling for ABM

Tang, D. and N. Malleson (2022). Data assimilation with agent-based models using Markov chain sampling. Open Research Europe 2(70). DOI: 10.12688/openreseurope.14800.1

Define an ABM using a particular scheme (similar to normal ABM definition)

New algorithm to allow efficient sampling from the space of all possible ABM trajectories

Use MCMC to combine the model with data to create a posterior

Monte-Carlo Markov Chain

Posterior estimates of predators and prey

Are they worth it?

Can urban digital twins help? Or should we focus on lower-hanging fruit?

Who will use them?

Will they help with ...

Life expectancy (can vary by 10 years)

Food banks

Pitiful rape (and others) prosecution and conviction rate

Obesity

Opportunities with connected data

Glasses in Classes

Connect education and health data to find children who had an ophthalmic deficit but had not had a followup appointment.

Evidence for cause of lower reading skills.

Glasses provided to schools

Undiagnosed Autism

Long waiting lists for autism assessments

But routinely-collected educational data give good predictions

Opportunity to speed up formal assessments in a data-driven way

Summary

1. Digital twins (of urban systems)

2. The role of agent-based modelling

3. Challenges for Urban Digital Twins

4. Progress towards Urban DTs

Data, model synthesis, emulation, real-time modelling

5. Are they worth it?

Can they deliver sustainable levelling-up?

Save the Date: GIScience 2023 (Leeds)

Short paper (5 pages) submission deadline

Mon 8th May 2023

Early-bird registration deadline:

Fri 14th July 2023

Conference:

13-15 September 2023 (workshops on Tuesday)

giscience.org/