PlaceAgen s: Mul i-S op Pedes ian I ine a ies as
Pla ial Flows on U ban Ne wo ks
James Williams[0000−0002−6199−4980]
Cen e o Dis up i e E a S udies,
Bi mingham Newman Uni e si y,
Bi mingham, UK
[email p o ec ed]
Abs ac . U ban li e is made by he places ha people me ge oge he
in o o mal ou ines. This manusc ip in oduces PlaceAgen s, a ame-
wo k o modelling mul i-s op pedes ian i ine a ies on ci y ne wo ks us-
ing open and ep oducible da a. The app oach uses he i ine a y as he
base beha iou al uni and links h ee co e p inciples: (1) access o open
da a sou ces, (2) enabling explici assump ions abou agen s, and (3) he
audi abili y o ou pu s. The app oach is suppo ed h ough a pipeline
o da a acquisi ion, agg ega ion, ou ing, and simula ion. A pedes ian
ou ing g aph and places a e ex ac ed om OpenS ee Map and spa ial
con ex is summa ised using he H3 g id. Resea che -in o med agen s
a e capable o planning sho sequences o place-based isi s, wi h mo e-
men ollowing a anspa en leng h-weigh ed A* cos on he ne wo k. In
a case s udy on No ingham and Bi mingham ci y cen es, he app oach
e eals high- h oughpu co ido s wi h a pla ial lens and he in luence
o i ine a y s uc u e on lows ac oss clus e s o Poin s o In e es . This
a icle p esen s he amewo k capabili ies in addi ion o he limi a ions
o simpli ied beha iou and he spa se a ailabili y o capaci y o opening-
hou s a ibu es. Finally, ex ensions o calib a ed choice models, expo-
su e accoun ing, and mo e c oss-ci y compa isons a e discussed.
Keywo ds: Agen -Based Modelling ·OpenS ee Map ·H3 ·Pla ial In-
o ma ion ·Pedes ian Mo emen ·U ban Analy ics
1 In oduc ion
Pedes ian lows in u ban spaces such as ci ies eme ge om chained ac i i ies
as opposed o isola ed ips, people s i ch oge he sequences o places o e -
ands, wo k, leisu e, o eme gency [1]. The co ido s obse ed a e co-p oduced
by clus e s o ameni ies, s ee s uc u es, and loca ions as opposed o single
o igin-des ina ion (OD) pai s alone e.g., [8]. P io wo k has sough o de elop
a la ge-scale agen -based model o daily pedes ian a ic lows o he ci y o
Salzbu g gene a ing housands o indi idual daily ac i i y chains o agen s ep-
esen ing bo h esiden s and ou is s [12], p oducing ich disagg ega ed mobili y
pa e ns o e en i e days. Simila ly, wo k has explo ed he de elopmen o open
2 J. Williams
da a and models o syn he ic demand moni o ing in Be lin [25]. The wo k p e-
sen ed in his manusc ip a emp s o ea u ban pedes ian mo emen as a
se ies o linked lows, in e media e s ops, and he i ine a y logic ha connec s
hem, suppo ing simula ion design in u ban analy ics and planning.
Using exis ing challenges in agen modelling as he impe us, his manusc ip
p esen s PlaceAgen s, an i ine a y-cen ed agen -based amewo k ha p io i-
ises exis ing map-based da a and audi abili y o esul s. The pipeline sepa-
a es da a acquisi ion, spa ial agg ega ion, ou ing, and simula ion in o modu-
la , inspec -able s ages. Pedes ian ne wo ks and Poin s o In e es (POIs) a e
sou ced om OpenS ee Map [18] ia OSMnx [3], enabling ich use cases o be
demons a ed h ough he amewo k. The app oach hen summa ises spa ial
con ex in o a H3 g id [9], enabling a spa ial-pla ial unde s anding o be o med
[23]. Inspi ed by he Mesa-Geo lib a y [21], agen s plan sho sequences o place
isi s and mo e on he g aph using an in e p e able leng h-weigh ed sho es -
pa h cos implemen ed wi h Ne wo kX [10]. This is b oadly based on he concep
o pla ial in o ma ion: he basic uni is no abs ac space bu uzzy places [2],
wi h agen s na iga ing be ween hem.
Guided by hese aims, his manusc ip add esses h ee esea ch ques ions
mo i a ing he s uc u e o he amewo k, he modelling suppo ed, and he
e alua ion: (RQ1.) How can mul i-s op pedes ian i ine a ies edis ibu e lows
ac oss u ban ne wo ks when buil on open da a?, (RQ2.) How can he in e ac ion
o clus e ed places and ne wo k s uc u e in luence he o ma ion and concen a-
ion o pedes ian lows?, and (RQ3.) How can an i ine a y-cen ed simula ion
enable in e p e able low pa e ns sui able o compa a i e analysis?
This manusc ip makes h ee con ibu ions: (1) i o e s a ep oducible pipeline
ha links OpenS ee Map o H3 agg ega ion, in e p e able ou ing, and a dis-
c e e ime-ma ched simula o wi h audi able agen e en logs; (2) i models
mul i-s op pedes ian walking place-based i ine a ies o e u ban ne wo ks wi h
explici anspo a ion simula ion cos s; and (3) i de i es edge pedes ian load
om hese e en logs, suppo ing he iden i ica ion o agg ega e co ido s and
agen mobili y ajec o ies.
The emainde o his manusc ip is o ganised as ollows: Sec ion 2 p esen
he o e all design o he amewo k and app oach including he p ocessing, ag-
g ega ion, and ou ing. Sec ion 3 desc ibes he s udy design and p esen s he
case s udy esul s o No ingham and Bi mingham. Sec ion 4 demons a es he
in e ac i e iewe . The penul ima e Sec ion (5) discusses he indings o he s ud-
ies and discusses hem in he wide con ex o he li e a u e. Finally, Sec ion 6
p esen s he conclusions and u u e wo k.
PlaceAgen s: Modelling Mul i-S op Pedes ian I ine a ies 3
2 F amewo k
Fig. 1: H3 dis ibu ion isualisa ions o Bi mingham, showing (le ) POI coun
dis ibu ion, and ( igh ) ne wo k edge dis ibu ion.
Fig. 2: H3 dis ibu ion isualisa ions o No ingham, showing (le ) POI coun
dis ibu ion, and ( igh ) ne wo k edge dis ibu ion.
PlaceAgen s is a i ine a y-cen ed amewo k o simula ing pla ial lows in u -
ban ne wo ks. The amewo k ea s he ci y as linked places embedded in a
pedes ian g aph, sepa a ing ep esen a ion (e.g., places o med om POIs [24],
ne wo k g aphs, and H3 spa ial con ex ) om beha iou (e.g., i ine a y choice
and ime-based mo emen ). I enables audi able e en s eams ha agg ega e o
4 J. Williams
edge-le el load wi hou impu ing missing a ibu es om open da a. PlaceAgen s
inpu s a single scena io ile (yaml) o de e mine p ocessing cha ac e is ics.
2.1 P ocessing
P ocessing begins h ough he cons uc ion o a pedes ian ne wo k and ca a-
logue o places downloaded di ec ly om OpenS ee Map [18] o each ci y cen e
bounding box. The ou able ne wo k is ex ac ed using OSMnx [3] wi h a walking
il e ha includes oo ways, pa hs, pedes ian oads, s eps, and sha ed su aces
while excluding p ohibi ed a eas. Topology is simpli ied o collapse in e media e
deg ee-2 nodes ha ca y no u ns, p ese e a di ec ed Mul iDiG aph so pa al-
lel segmen s and one-way es ic ions emain explici , and compu e me ic edge
leng hs. Fo each edge, iden i ie uples, geome ies, highway class, access lags,
and leng h is me es is e ained. This yields a ou able pedes ian g aph ha
e lec s OpenS ee Map ags wi hou pe o ming addi ional impu a ion.
Using he same bounding box, POIs a e ex ac ed using OSMnx’s geome y
que y wi h a whi elis o OpenS ee Map ags co esponding wi h he scena ios
place ypes (e.g., ca es, lib a ies, supe ma ke s, es au an s, banks, he i age si es,
museums, and as ood). Polygons a e con e ed o ep esen a i e poin s (cen-
oids cons ained o lie wi hin he sou ce geome y), and key a ibu es such
as name, ags, opening hou s, capaci y, and he OSM iden i ie a e e ained.
To enable node- o-node ou ing, each place poin is snapped o i s nea es ne -
wo k node, wi h he snapped node id and snap dis ance eco ded in me es. No
syn he ic nodes o edges a e c ea ed and places wi hou a plausible snap ( o
example, beyond a h eshold) a e lagged and excluded om ou ing, any POI
wi h missing a ibu es is le missing.
2.2 Spa ial Agg ega ion
PlaceAgen s in eg a es a disc e e global g id sys em (DGGS) o segmen he
map in o dis inc , indexable cells [13]. Fo he wo k conduc ed in his a icle,
he H3 g id was selec ed due o i s ma u e unc ionali y [13], enabling a uni o m
ep esen a ion o spa ial con ex [9]. DGGS a e inc easingly used in exis ing li -
e a u e, such as classi ying walkable g id cells in leisu e ou e ecommenda ions
[22], es ima ing u ban a eas based on his o ical census da a [19], and spa ial ag-
g ega ion om SAR images [14]. The mul i- esolu ion hie a chical implemen a-
ion enables consis en and p edic able agg ega ion o ne wo k and place ea u es
ac oss mul iple scales, mo i a ing i s use in his wo k.
Ne wo k and place con ex is summa ised in o a H3 g id cell [9] as demon-
s a ed in Figu es 1 and 2 a a ixed esolu ion selec ed in he scena io. Fo each
scena io, a se o hexagon cells co e ing he ull ex en a e used wi h he H3
index a ached. POIs a e snapped o hexagons h ough associa ing he poin o
he polygon, wi h da a hen being summa ised by cell. To ela e he simula ed
mo emen o spa ial con ex he o al edge leng h is epo ed h ough in e sec -
ing geome ies and edge load in p opo ion o segmen o e lap. Fo mally, wi h H
PlaceAgen s: Modelling Mul i-S op Pedes ian I ine a ies 5
as a hexagon cell and Las he di ec ed edge load, place coun s, ne wo k leng h,
and dis ibu ed load a e epo ed as in Equa ion (1).
poiH=X
p∈H
1, LH=X
e
leng h(e∩H),¯
LH=X
e
Le
leng h(e∩H)
leng h(e).
(1)
Whe e poiHcoun s snapped places in hex H;LHsums me ic leng hs o edge
segmen s inside H;¯
LHdis ibu es each edge’s load in o hexes in p opo ion o
segmen leng h.
2.3 Agen s and Engine
Agen s ep esen he pedes ians execu ing sho i ine a ies o place isi s on
he simula ed ne wo k. Each agen main ains a cu en node, an o de ed a -
ge lis o places snapped o he ne wo k, a planned pa h, and a ini e s a e
o : {planning,mo ing,a i ed,s uck, inished}. O igin poin s a e sampled om
ne wo k nodes wi hin he scena io ex en and snapped o he ne wo k and a -
ge s a e sampled om he places ex ac ed il e ed om he scena io’s ag il-
e . Whe e he OpenS ee Map da a has opening hou s, places a e ea ed as
empo a ily no possible. Whe e opening hou s a e missing, he isi p oceeds
wi hou a empo al limi a ion. This i ine a y-cen ed design ollows wo k ha
o eg ounds wo k wi h u ban subdi ision in pedes ian mo emen simula ions
[6].
Each leg o he ou e is compu ed on he di ec ed pedes ian Mul iDiG aph
using he A* g aph a e sal algo i hm wi h he edge leng h as objec i e and a
s aigh -line heu is ic, ollowing Equa ion 2, implemen ed wi h Ne wo kX [10][11].
Du ing hese phases he agen is conside ed o be planning o mo ing, i no ea-
sible pa h is possible he agen ansi ions o he s uck s a e and e mina es.
Agen s mo e along he ne wo k by a ixed dis ance pe ick, upon eaching a
des ina ion node he a i ed s a e is emi ed and he agen e u ns o he plan-
ning o inished s a e. Du ing his p ocess he di ec ed edge load is calcula ed
as p esen ed in Equa ion 3, and is also summa ised in o H3 cells (Equa ion 1).
(n) = g(n) + h(n), g(n) = X
e∈pa h(s→n)
leng he, h(n) = dgc(n, goal).
(2)
Whe e sis he s a node, leng heis he edge leng h, and dgc is g ea -ci cle
dis ance; wi h leng h weigh s, his admissible.
2.4 Ou pu s
PlaceAgen s p oduces a se o a e ac s om which all igu es, s a is ics, and
simula ions a e compu ed. The pe - ick e en log eco ds e e y mo e and a i al
e en , om which i is possible o de i e edge loads as coun s o a e sals as
6 J. Williams
demons a ed in Equa ion 3, which a e hen a ached o he pedes ian ne wo k
o mapping and analysis. H3 summa ies a e also gene a ed a he scena ios ixed
esolu ion: place coun s, ne wo k leng h, and load based on segmen o e lap.
Finally, a un summa y is p oduced which cap u es he scena io, ne wo k, and
agg ega e ou comes (bbox, du a ion, numbe o agen s, e c).
Repo ed based on hese ou pu a e ac s a e: (1) comple ion me ics based
on he sha e o agen s comple ing all s ops, and he mean s ops comple ed; (2)
pa h s a is ics based on he mean, media, and deciles o leg leng hs based om
he A* (Equa ion 2); and (3) plausibili y checks including he spea man ank
co ela ions be ween node/edge load and cen ali ies (deg ee, edge be weenness)
[20][7] and he Gini coe icien o he edge-load dis ibu ion o analysing co ido
concen a ion [5]. All measu es a e compu ed di ec ly om he open-da a de i ed
ne wo k and expo ed.
Le=X
a
X
k
1{a a e ses edge ea k}.(3)
The indica o 1{·} equals one when agen amo es along edge ea ick k, else
ze o; Leis a aw coun sui able o mapping and summa ies.
3 Case S udies
The PlaceAgen model is e alua ed in wo UK ci y cen es: No ingham (a ound
he Old Ma ke Squa e and Lace Ma ke a eas) and Bi mingham (a ound New
S ee S a ion, he Bull ing, and Vic o ia Squa e). Bo h simula ions use he same
pipelines, H3 esolu ion, ime s eps, and POI ag selec ion. The only a ia ions
a e he local ne wo k and POI mix di e ing be ween uns. Scena io pa ame e s
a e p esen ed in Table 1 demons a ing he simila i ies o he simula ions and
ci y-speci ic cha ac e is ics a e p esen ed in Table 2, highligh ing he compa a-
i ely la ge scope o Bi mingham pa icula ly in nodes, edges and POIs.
Table 1: PlaceAgen s scena io pa ame e s de ined in he scena io se ing ile
p esen ed by ci y.
Ci y
BBox
(min lon, min la ,
max lon, max la ) Agen s Du a ion S ep H3 size POI ags
No ingham
[-1.164, 52.948,
-1.131, 52.967] 1000 360 5 9 see no e
Bi mingham
[-1.918, 52.472,
-1.885, 52.492] 1000 360 5 9 see no e
No e: Tags used: ca e,pa k,lib a y,supe ma ke , as _ ood, es au an ,
pha macy,bank,ba ,pub.
PlaceAgen s: Modelling Mul i-S op Pedes ian I ine a ies 7
Table 2: Da a cha ac e is ics o Bi mingham and No ingham ex ac ed om
he inpu da a om OpenS ee Map [18].
Ci y Nodes Edges POIs Hexagons Median edge leng h [m]
No ingham 2460 6446 1317 26 19.466
Bi mingham 3297 8918 1858 54 24.475
(a) No ingham (b) Bi mingham
Fig. 3: Dis ibu ion o pa h leng hs (nodes) o he simula ed i ine a y legs o
No ingham and Bi mingham.
The case s udies bo h comple ed success ully, wi h 100% o agen s eaching all
in ended des ina ions on he simula ed ne wo k. Figu e 3 p esen s a compa ison
o he pa h leng hs o bo h No ingham (a) and Bi mingham (b), indica ing a
sligh ly longe a e age edge leng h in he Bi mingham scena io. The emainde
o his sec ion will p esen he esul s o bo h ci ies and p o ide desc ip i e
s a is ics ega ding he simula ions.
3.1 No ingham Resul s
The No ingham simula ion esul ed in an in e es ing se ies o esul s ep esen -
ing mo e concen a ed co ido s h ough he co e s ee g id, wi h ligh e use on
backs ee s and side oads. Agen s in he No ingham simula ion had a mean
s op coun o 4, ep esen ing a dis ibu ed amoun o i ine a y simula ions. Fig-
u e 4 demons a es sample agen ajec o ies (a) and hese condensed edge loads
in p ac ice (b). The esul ing leg leng hs a e compac o a ci y-cen e i ine a y
median = 631 m; 10-90% = 263-1,235 m), while load inequali y is high (Gini
= 0.82), indica ing a small se o dominan segmen s. The plausibili y checks
show a s ong associa ion be ween node load and node be weenness (Spea man
ρ= 0.83) and a mode a e associa ion wi h deg ee (ρ= 0.52). The hexagon le el
load is weakly co ela ed o No ingham ( = 0.16).
The No ingham esul s ep esen mo e concen a ed pedes ian pa e ns
which add ess RQ1 by showing how sho , mul i-s op i ine a ies can gene a e
highly unequal pedes ian lows wi hin a compac ci y-cen e scena io, e en when
using simpli ied beha iou al assump ions.
8 J. Williams
(a) No ingham sample agen ajec o ies. (b) No ingham edge load.
Fig. 4: Demons a ion o sample agen i s -leg ajec o ies o No ingham (a)
and edge load dis ibu ion (b) o he same a ea.
3.2 Bi mingham Resul s
The Bi mingham simula ion demons a es compa a i ely longe legs (median
= 1,243 m; 10-90% = 441-2,131 m) and lowe concen a ion han No ingham
(Gini = 0.77), e lec ing he la ge g aph and wide ameni y sp ead. All agen
inished wi h a mean o ou s ops. The simula ion esul s p esen ed in Figu e 5
shows s ong plausibili y checks, node load co ela es e y s ong wi h node
be weenness (Spea man ρ= 0.87) and mode a ely wi h deg ee (ρ= 0.57). The
hexagon le el load co ela es a a mode a e le el oo ( = 0.52), indica ing ha
ameni y clus e s align wi h high h oughpu co ido s mo e clea ly han he
No ingham esul s.
The Bi mingham case s udy p esen s wide and less concen a ed lows which
ep esen he la ge ne wo k ex en and numbe o places applicable o he lo-
ca ion. This case s udy add esses RQ2 by demons a ing how la ge ameni y
clus e s and ne wo ks in e ac o shape pedes ian low and co ido o ma ion.
PlaceAgen s: Modelling Mul i-S op Pedes ian I ine a ies 9
(a) Bi mingham sample agen ajec o ies. (b) Bi mingham edge load.
Fig. 5: Demons a ion o sample agen i s -leg ajec o ies o Bi mingham (a)
and edge load dis ibu ion (b) o he same a ea.
Combined wi h he No ingham case s udy, he esul s demons a e how a con-
sis en i ine a y design can p oduce di e en mo emen s uc u es depending on
he con ex o he ci y.
4 In e ac i e Viewe
Figu e 6 p esen s a snapsho o he gene a ed PlaceAgen s iewe o No ingham
Ci y Cen e, p esen ed on a OpenS ee Map basemap [18]. Pedes ian edges a e
isualised by simula ed load, places appea as la ge ci cula o ange poin s, agen s
ende as anima ed mo ing ma ke s du ing playback, and can lea e a colou ed
ace on he map. The o e laying le panel con ols laye s, ci y selec ion, and he
sampling o agen s displayed, and he bo om ba enables he use inspec ing o
sc ub h ough he simula ion playback. In his No ingham Ci y Cen e scena io
1,000 agen s a e simula ed, wi h each segmen changing he isualised colou
o he agen . The iewe enables he illus a ion o how i ine a ies begin o
concen a e low along a small se o co ido s while backs ee s and subu ban
oads emain ligh ly used.
5 Discussion
5.1 In e p e a ion o Resul s
The i ine a y pe spec i e adds addi ional explana o y de ail beyond OD ap-
p oaches. Ac oss bo h cen es i is possible o obse e unequal co ido o -
ma ion om sho mul i-s op chains, wi h high concen a ion in No ingham
