Bridging agent-based modelling and R with nlrx: simulating pedestrian’s long-term exposure to �air pollution

Hyesop Shin

Ph.D Candidate

Dept. of Geography and Fitzwilliam College

University of Cambridge

@hyesop

TOC

Objective

Agent-based modelling

R Codes

Outcomes

Summary

Files, Codes and Tutorials:

http://tiny.cc/nlrx

Project Objective

Project objective

• This study aims to estimate pedestrian’s exposure to acute air pollution in Seoul districts using agent-based simulation
• How does socioeconomic group potentially affect health outcomes?
• How could health levels change under different pollution scenarios?

Agent-based modelling

Agent-based modelling (ABM) is a computational method that focuses on individual’s movements and interactions that can affect the system structure�e.g. Social media, epidemics, decision-making

Components

• Agents = turtles
• Environments = patches

ABM setup

How ABM works: NetLogo example

Agent-based modelling: Settings

​

Study area + demographic info

Agent setup and create destination

Measure health loss and recovery

Export file as export.csv

Import files to R

96,822 patches

Agent-based modelling: Settings

​

Study area + demographic info

Agent setup and create destination

Measure health loss and recovery

Export file as export.csv

Import files to R

• The population of Gangnam is allocated into three groups: �Under 15, 15-64, Over 65
• A day is splitted into two time sequences: �Work hours (09-19 hrs), Home (20-08 hrs)
• Agents have no previous exposure experienced
• Agents aged 15-64 follow OD matrix while restricting other groups' movement range close to their origin

Agent-based modelling: Settings

​

Study area + demographic info

Agent setup and create destination

Measure health loss and recovery

Export file as export.csv

Import files to R

If PM₁₀ >100

Agent-based modelling: Settings

​

Study area + demographic info

Agent setup and create destination

Measure health loss and recovery

Export file as export.csv

Import files to R

1.2 hours �for a single run

Low Quality�Images

Exported as (uncleaned) .csv

50 iterations

You will end up doing this...

Time-consuming!

Solving current problems from NetLogo

How can R reduce human intervention during analysis?

How can R improve the quality of figures?

How can R increase the simulation's clock speed?

01

​

02

​

03

ABM with nlrx: �a tool to setup and execute NetLogo simulations from R�Netlogo + R + XML

Stage 1: Install packages nlrx

Sys.setenv(JAVA_HOME= ‘XXXX‘) #Varies by OS

library(nlrx)

library(tidyverse)

library(rcartocolor)

library(ggthemes)

​

nlpath <- file.path("/home/hs621/NetLogo 6.0.4")

modelpath <- file.path(path,"Gangnam.nlogo")

outpath <- file.path("/home/hs621/out")

​

nl <- nl(nlversion = "6.0.4",

nlpath = nlpath,

modelpath = modelpath,

jvmmem = 1024)

Java setup

Load pkgs

Assign path

Create an nl object

Codes and tutorials: http://tiny.cc/nlrx

Stage 2: Attach an experiment (1/2)

nl@experiment <- experiment(expname = "seoul",

outpath = outpath,

repetition = 1,

tickmetrics = "true",

idsetup = "setup",

idgo = "go",

runtime = 8764,

evalticks=seq(1,8764, by = 100),

Codes and tutorials: http://tiny.cc/nlrx

Stage 3: Attach an experiment (2/2)

constants = list("PM10-parameters" = 100,

"Scenario" = "\"BAU\"",

"scenario-percent" =

"\"inc-sce\""),

variables = list('AC' = list(values=c(100,150,200))),

metrics.turtles = list("people" = c("xcor", "ycor",

"homename", "destinationName", "age",

"health")

metrics.patches = list("patch" = c("pxcor", "pycor", "pcolor"))

Codes and tutorials: http://tiny.cc/nlrx

Stage 4: Attach a simulation design and run model

nl@simdesign <- simdesign_distinct(nl=nl,

nseeds=1)

​

init <- Sys.time()

results <- run_nl_all(nl = nl)

Sys.time() - init

setsim(nl, "simoutput") <- results

write_simoutput(nl)

Iteration sampling

Run

Add results to nl list

Codes and tutorials: http://tiny.cc/nlrx

Stage 5: Submit batch jobs to reduce time

Post-simulation

Result structure: nested tibble

General info: Iteration, Scenario, Random seed, Step,

Agent info: Count turtles, Breed, X, Y, home, destination, age, health

Patch info: X, Y, patch colour

⇒ unnested� result

Mapping unhealthy population with ggplot2

Codes and tutorials: http://tiny.cc/nlrx

ggplot() +

facet_wrap(~`[step]`, ncol= 10) +

coord_equal() +

geom_tile(data=patches, aes(x=pxcor, y=pycor, fill=pcolor), alpha = .2) +

geom_point(data=turtles, aes(x = pxcor, y = pycor, color = age), size=1) +

scale_fill_gradient(low = "white", high = "grey20") +

scale_color_manual(breaks=c("young", "active", "old"),

values = c("young" = "#56B4E9",

"active" = "#E69F00",

"old" = "#999999")) +

guides(fill=guide_legend(title="PM10")) +

ggtitle("Unhealthly Population after a long-term exposure") +

theme_minimal() +

theme(axis.line=element_blank(),

axis.text.x=element_blank(),

axis.text.y=element_blank(),

axis.ticks=element_blank())

​

Density plot with ggplot2 & �Animations with gganmiate

Codes and tutorials: https://tiny.cc/nlrx

gridextra & directlabels for HQ images

Codes and tutorials: http://tiny.cc/nlrx

grid.arrange()

direct.label()

Summary and Contribution

Summary

• Contents
• Disparities in health outcomes are likely to depend on demographic status
• When the vulnerable group (the old and young) is exposed over a long period, road proximity causes additional health degradation
• R & nlrx
• Workload: 60% on NetLogo, 30% on R, 10% HPC
• nlrx works as a compiler
• nlrx results in a tibble format that can be plotted in a variety of figures for different purposes with ggplot, gganimate

Contributions from nlrx

• Time saving
• Fast iterative process
• The simulation ends with only a paragraph of codes
• Solves fat finger issues
• No need to delete readme text from the NetLogo output
• Convenience
• Doesn’t need rJava installation (Really helps when running HPC)
• Silent machines
• Code categorical variables

For more information...

https://tiny.cc/nlrx

Paper

Tutorial

Thank you!

@hyesop

hs621@cam.ac.uk

@mrsensible

Agent-based modelling: advantages

• Is able to simulate human-environment interaction
• Can analyse adverse health impact by social groups
• Follows trajectories of individuals over time
• Envisages effects from possible scenarios ("What if…?")
• NetLogo is the most widely used software in the ABM world

Air Pollution in South Korea (March, 2019)

Songdo, Incheon (Mar.5th, 2019)

Central Seoul (Mar.6th, 2019)

PM_2.5 > 200

PM_2.5 > 140

Urbanisation and Air Pollution

NO₂ trend in a decade (2005-2015), Satellite imagery taken by NASA, 2015

Urbanisation and Air Pollution

NO₂ change in East Asian countries

Traffic in Seoul CBD

Health threat may differ by demographic factors

• Children are inherently more susceptible to air pollution as their lung function and immunological systems are still developing (Pearce et al., 2006)
• Higher risks due to the PM₁₀ exposure were observed for elderly individuals - COPD, stroke, etc (Halonen et al., 2016; Wang et al., 2016)

More attention should be given to how travel behaviours differ by social groups (e.g. age), and how health loss are manifested in each group after a long-term pollution exposure

Stage 1: Install and finish coding in NetLogo

• Install NetLogo >= 5.3, NetLogo 6 is preferred
• Java required

​

​

Agent-based modelling: advantages

• Is able to simulate human-environment interaction
• Can analyse adverse health impact by social groups
• Follows trajectories of individuals over time
• Envisages effects from possible scenarios ("What if…?")
• NetLogo is the most widely used software in the ABM world

NetLogo Example

96822 patches