Read more

• Levinson, D., Sarkar, S., Tajik, N. and Ermagun, A. (2026) Access-based network growth: preferential attachment with capacity, cost, and connectivity. Applied Network Science, October 23, 2026. [doi]

This study proposes a reduced-form, access-based model of network growth in which attachment depends jointly on accumulated connectivity, residual capacity, and origin-destination pair-specific link cost. The kernel contains Barabási-Albert (BA) preferential attachment, spatial attachment, and capacity-limited growth as limiting cases. At N = 1000, α = 1, and κ = 2, the capacity-only specification (β = 1, K = 16) reduces mean maximum degree from 82.31 in the BA limiting case to 16.00, while the spatial-only specification (φ = 1) reduces mean edge length from 0.521 to 0.351. Increasing φ from 0 to 2 shortens mean edge length from 0.526 to 0.157, and increasing κ from 1 to 4 raises clustering and reduces average path length. With mean capacity near 16, integer-uniform and rounded-lognormal capacities raise mean maximum degree to 23.06 and 49.00, respectively. Crossing-handling and access-guided extensions represent junction formation and compare four configured transport-growth strategies using a common ex post access measure. The framework yields directional hypotheses for empirical calibration.

Data/code: https://github.com/dlevinson/attachmentlab-js/tree/main/reproducibility/access-based-network-growth

• Levinson, D. M. (2026) Estimating Pedestrian Flows and Signal Delay from Sydney Traffic-Signal Event Histories. Findings, October 20, 2026. [doi]

Traffic-signal histories record first pedestrian calls but omit subsequent arrivals. We apply a calibrated first-actuation estimator to Sydney histories for 18 September 2018. Screening retains 5,114 movement-hours at 174 sites, yielding an estimated 331,729 crossings. Estimated delay is 3,671.5 pedestrian-hours assuming full compliance and 2,746.3 hours under a transferred behaviour scenario: 39.8 and 29.8 seconds per crossing. Median site flow is 66.5 crossings per usable hour.

Keywords:

pedestrian flow; pedestrian delay; traffic-signal histories; pedestrian actuation; censored estimation

Data/code: https://github.com/dlevinson/paper-2026-25-estimating-pedestrian-flows-and-signal-delay-from-sydney-trafficsignal-event-histories

Figure 2. Estimated pedestrian crossings at site 516. Panel A shows the 24-hour profile for 18 September 2018. Crosses mark hours that are not estimable under the screening rules; the hatched 22:00 bar uses three of the four pedestrian movements. Panel B compares seven consecutive days using identical screened movement and clock-hour coverage; the Tuesday analysis date is highlighted. Panel C shows monthly typical-day estimates from September 2017 through August 2018 using the same movements and clock hours in every month. Orange labels identify months in which one incomplete date is excluded. Monthly values are typical-day estimates, not monthly totals.

I believe this logic holds in 2026. Since then, the bus proposal has been evaluated for accessibility and we have produced cleaner maps, renumbered the LRT proposal,1 along with other minor updates.2 I will trust readers of this Blog already understand why buses are important, and how rapid bus networks should be designed.

Light Rail Network

FAST combines a broadly east–west and north–south rapid and frequent bus grid with light-rail extensions and new surface corridors. The buses are more numerous, and will carry more people more kilometres, but on high use corridors, LRT still matters, and serves flows and distances between those best served by buses and those best served by Trains and Metros.

Figure 1. Sydney light rail: operating L1–L3, their proposed extensions, L6–L8 and the Elizabeth Street option.

Figure 1 shows the proposed LRT network for Eastern Sydney. The L1 extension continues south from Central through Green Square towards Mascot. L2 extends west along Broadway, City Road, King Street and Princes Highway towards Wolli Creek, and east from Randwick to Coogee. L3 extends west through Broadway, Parramatta Road, Norton Street and Marion Street towards Five Dock, and south from Kingsford through Maroubra towards Little Bay.

So in short, frequency would increase on the parts of the L2 and L3 east of Central, so additional LRT vehicles could be routed on the extensions, and some of the vehicles from the extensions would be routed into the city. The Elizabeth Street option is presented as with high frequency services George Street operation may need additional capacity. There could be one long line from Five Dock to Central up George Street to Circular Quay back down Elizabeth Street to Central to Little Bay, and similarly from Wolli Creek to Coogee). Whether these deserve new designations, I will leave to marketing, but probably yes. The map suffixes “sx”, “wx” and “ex” identify southern, western and eastern extensions.

L6, links West Ryde, Ryde, Gladesville, Drummoyne and Rozelle with the city, Oxford Street, Bondi Junction and Bondi Beach. It would use part of the L1 corridor. Shared-track capacity, junctions and the city connection require operational and engineering assessment.

L7 and L8, share a Wynyard–Harbour Bridge–North Sydney approach before serving Northbridge and Lane Cove. With Metro operating, these lines need a clear role in local access. Bridge space, gradients, street capacity, interchange arrangements and a bus alternative must be tested before selecting light rail.

In Parramatta (Figure 2), FAST extends the network from North Parramatta through Baulkham Hills to Castle Hill (L4-C). L4-E would close the Carlingford–Epping gap. A southern alignment (OP-A) is drawn and should remain in the mix, serving different markets than the L5.

Figure 2. Parramatta: operating L4, the government’s L5 corridor, and the proposed L4-C, L4-E and OP-A routes. Diamonds identify areas. L4 and L5 share the approach from Parramatta Square to Camellia, then cross the river separately towards Yallamundi and Rydalmere.

Rapid Bus Network

While Sydney’s coastline, rivers, hills and street pattern make a perfect grid impossible, FAST uses available streets and connects to rail stations and important destinations. The bus network identifies routes by their alignment direction. H identifies generally east–west corridors, with even-numbered intermediate routes; V identifies generally north–south corridors, with odd-numbered intermediate routes.

The diagrams simplify bends and distances and show selected named places. Solid light-rail lines are operating corridors; short dashes mark FAST concepts and longer dashes the government’s Parramatta Light Rail Stage 2.

Orange nodes mark Train transfers and teal nodes Metro transfers. An orange ring with a teal centre marks both. Dashed outlines identify future rail connections. These are transfer opportunities in the named area; final stops, walking links, gradients and platform arrangements still need design.

Figure 3. The 12 proposed H bus corridors in blue.

Figure 4. The 22 proposed V bus corridors in purple.

Figure 5. Rapid buses: Western Sydney and the Inner West.

Figure 6. Rapid buses: Inner Sydney, the east and Northern Beaches

Accessibility analysis

Nixon Huynh’s November 2025 honours thesis, Network Evaluation of a Bus Rapid Transit Network for Sydney, compares FAST with the Committee for Sydney’s Plan B: Better Buses for Sydney.3 It uses r5r to estimate employment access within 30, 45 and 60 minutes, including walking, waiting and transfers. Results are weighted by residential population. Huynh 2025, pp. 17–23.

Fourteen of the fifteen highest-ranked individual routes come from FAST; the input pool contains 34 FAST and 20 Plan B routes. H64, V23, V13 and V21 lead the ranking, followed by Plan B’s CBD–Bondi corridor, CS6 (which FAST proposes as part of the L6 line). These results support investigating short connections and interchange quality alongside long arterial routes.

Read more

How the Sydney cap works

New South Wales introduced its toll cap in January 2024. Drivers pay the full toll when they travel, and get some returned later as a rebate.

• For each e-tag or number plate, Service NSW (an agency, but to most people an app on their phone) refunds weekly (Monday to Sunday) toll spending above a threshold. The threshold was $60 and is $50 until July 2027. Refunds are limited to $350 a week and $5,000 a year.

• Only vehicles registered for private use qualify. [i.e. no commercial vehicles]

• Drivers link their toll account to a MyServiceNSW account and lodge a claim each quarter. Refunds are not automatic.

By mid-2025 the scheme had returned $139 million to 500,000 motorists, mostly in western Sydney, and another $100 million set aside remains unclaimed. NSW has since made the cap effectively permanent. Under the August 2026 deal with the motorway owners, the operators contribute $75 million over five years, and the state pays the rest, which they probably get back with induced demand from the lower toll prices.1

What Victoria is being offered

Both parties would copy the Sydney model on CityLink, EastLink and the West Gate Tunnel, with claims lodged through Service Victoria.

• Labor: $30 a week, for 2027 only. Cars, utes, vans and motorbikes qualify, including up to five vehicles per small business. Up to $5,000 per vehicle. Labor says about 200,000 account holders would benefit, at a cost of $191 million. It would also cut the $16 fee for paper toll notices to $5.

• Coalition: $50 a week for two years from 1 July 2027. Up to $6,000 per vehicle over the two years. The Coalition expects about 100,000 beneficiaries, at a cost of $180 million. Trucks, rideshare vehicles, and fleets of businesses with turnover above $10 million are excluded.

Who pays

The cap subsidises drivers (transponder owners or vehicle owners) with public funds. Transurban (which owns CityLink and the West Gate Tunnel) and the owners of EastLink still receive the full toll. Transurban is owned by its shareholders, and much of that is superannuation funds. If the cap brings more trips onto the motorways, the owners earn more.

The rebate comes from general revenue, so it is paid in part by people who don’t drive, people who drive but don’t use toll roads, and people who can’t afford a car.

A subsidy to driving

Most transport policy aims to reduce vehicle travel, to cut congestion, pollution and crash risk. The toll cap subsidises driving and will increase it overall. Once a driver passes the weekly threshold, additional tolled trips cost them nothing, up to the weekly limit.

Some trips will move from surface streets onto motorways. That is beneficial, since motorways are safer per kilometre and keep through traffic off local streets, but it does not offset the increase in total travel. We have started research on how much additional travel the NSW cap generated, and it is too early to draw conclusions.

It is worth noting there is a similar $50/week cap on public transport fares in Sydney, and a similar cap is being proposed in Victoria. This is different, because we generally, as policy, want to induce public transport use rather than car use. It is not optimal though, as our work on the economic effects shows, and reduces revenue to the transit operator.

• Mann, I. and Levinson, D. M. (2026) Fare enough: A quantitative spatial transport model for transit pricing. Transportation Research Part A: Policy and Practice, Volume 214, December 2026, 105237. [doi]

Salience

Victorians would need to remember to claim back through Service Victoria every quarter, which is more (electronic) paperwork to file. The toll operators already know what each tag spends each week, so the rebate could be automatic, but it isn’t, which saves the government (and public) money, permitting politicians to pander to their prospective partisans.

A driver who isn’t paying close attention quarterly might drive more thinking they are capped, and then never claim. In Sydney, $100 million set aside for the scheme was still unclaimed in mid-2025, which suggests many eligible drivers did not claim.

How much behaviour changes depends on how much a driver already spends:

• Infrequent users, spending under $30 a week, won’t be affected.

• Heavy regular users, mostly commuters already over the threshold, get a kickback but probably won’t change much.

• Drivers near the threshold are the most likely to change, adding freeway trips they would otherwise have skipped or taken on another route.

A better fix

In 2025 I am quoted in the Sydney Morning Herald that removing the NSW cap without a replacement would hit low-income households in western Sydney hardest, because they pay a lot in tolls and have few alternatives. I still think so. But the rebate is a patch on a toll system that is regressive to begin with.

A better fix is to change the tolls themselves, through the unified distance-based tolling, such as recommended by the Fels–Cousins Independent Toll Review, which public ownership of the motorways would make possible. (And of course to lower prices in the off-peak and in uncongested areas).

NSW’s August 2026 deal cut tolls modestly on a few roads and left WestConnex, the largest network, out of the price cuts.

Victoria has three toll roads, and North East Link, due in 2028, will be tolled by a state-owned company. That leaves Victoria better placed than NSW to move to consistent, distance-based tolls instead of paying weekly rebates.

Media links:

• Labor pledges $30 toll cap as Coalition targets gas ban (ABC, 18 October 2026)

• A toll order: Labor and Coalition fight it out over road toll relief for Victorian drivers (Herald Sun, 18 October 2026) [gift link]

• Toll price reductions set for multiple Sydney motorways including Lane Cove Tunnel and M2 (ABC, 3 August 2026)

• Josh’s toll bill is about to rise 233 per cent. He’s not alone (Sydney Morning Herald, 10 July 2025)

Read more

• Askari, S., Tilahun, N. and Levinson, D. (2026) Determinants of automated speed camera location choice. Journal of Transport Geography, Volume 137, December 2026, 104869. [doi]

Although speed cameras are widely documented to reduce speeds and improve road safety, less is known about how jurisdictions decide where to place them and whether placement aligns with stated safety and equity goals. This study examines speed camera location decisions in Chicago. We compare selected camera sites with legally eligible but non-selected candidate sites within school and park safety zones, and analyze how crash history, nearby child population share, neighborhood socio-demographics, political engagement, and roadway-user composition are associated with camera placement. Models are estimated using the full eligible-site sample, a propensity-score matched sample, and a stricter caliper-matched sample. The results show that crash history and nearby child population share are the strongest predictors of camera installation. After accounting for these safety-related criteria, camera placement is negatively associated with neighborhood poverty and population density. Tract racial composition is not a robust predictor of camera placement, while roadway-user composition and voter turnout show less stable associations across specifications. Overall, the findings suggest that Chicago’s initial speed camera deployment was largely aligned with stated safety criteria. However, the negative association between poverty and camera placement points to a safety-equity tradeoff. Municipalities should ensure that efforts to reduce fine burdens in disadvantaged areas do not leave high-risk locations without other safety interventions.

Keywords:

Automated traffic enforcement, Speed camera, Camera site selection, Propensity score matching, Safety

• Ji, A., Ma, Z., Yang, H., Khattak, A. J. and Levinson, D. (2026) Learning from common knowledge: A multi-agent cooperative ramp merging framework for decentralized autonomous vehicles. Transportation Research Part E: Logistics and Transportation Review, 105275. Available online 12 October 2026 (in press, corrected proof). [doi]

This paper proposes a multi-agent reinforcement learning (MARL) framework that combines an offline shared pool of information (SPI) with hypernetworks to enable cooperative on-ramp merging among high-level autonomous vehicles (AVs), leveraging common knowledge for spontaneous cooperation. First, we discuss the construction method of the common knowledge base-SPI from two approaches: model- and data-driven methods. Then, we design an on-board fine-tuning algorithm that enables each agent to refine its policy from local observations without any communication protocols. Based on the above advances, we develop a hybrid SPI-based on-ramp merging strategy driven by the physical model and human driving data. Finally, the proposed framework is evaluated in simulated peak- and off-peak traffic scenarios calibrated by the Zen trajectory dataset. The results demonstrate improved traffic efficiency outcomes and fewer collision occurrences than several baselines under the evaluated conditions. The findings reveal possible research directions for future work and provide insights for the application of the decentralized-training-decentralized-execution framework in autonomous driving.

Keywords:

Autonomous vehicles; Multi-agent reinforcement learning; Ramp merging; Shared information pool

As a show about representation, it was quite progressive for the time, with speaking roles for disabled and black characters, and on occasion showing the barriers they face in modern society.

It has the typical Raymond Burr issue of the overbearing protagonist, speaking as a Perry Mason fan. The wheelchair has not humbled him. It is not so much that he is smart, but that he has unique information that no one else could possibly have.

The show also has the typical TV police procedural problem of “this time it’s personal”. Far too many crimes involved the officers’ friends and family for comfort. The key lesson is don’t befriend a cop.

Ironside addressed pressing issues of the day like Racism and Vietnam and Drugs and Hippies and Defendants Rights and University student gun clubs (with clock tower shootings alluding to recent events in Texas) and Back Alley Abortionists and Rioters and Nazis. There were only 30-40 murders per year in San Francisco in the late 1960s, so there was more crime on the show than in reality. Ironside is not as bad as Longmire in this regard, but clearly gives TV viewers a distorted picture of criminal violence, when most of the crime is petty theft (as I personally observed in 2024). It is thus complicit in the collapse of trust.

Robert T. Ironside was Chief of Detectives before being shot at the opening of every episode to a great theme by Quincy Jones, forcing him to use a wheelchair. His new role as Special Consultant let him work whatever cases he wanted. Mark Sanger is Ironside’s aide, though not an officer, who Ironside rescued from the streets and a life of crime. Sanger is attending law school. Officer Eve Whitfield (a woman cop too, adding to representation), and Sgt Ed Brown (as white and male as they come) are his police staff. Ironside works from home, which is to say, strangely, he lives at the police station and has an office in his “living room”. He rides around San Francisco in a converted police paddy wagon (and later a van, as I assume an automaker paid for placement), driven by Mark when he is not pushing the wheelchair.

The University of California, Berkeley sometimes serves as their advanced crime lab in a precursor to CSI.

The show also has guest appearances of actors who later become famous like Harrison Ford, David Carridine, Gavin MacLeod, Ricardo Montalban, Martin Sheen, and Gary Collins among others, and some who had some experience like Bill Bixby, Burgess Meredith, Robert Alda or Anne Bancroft.

Relations with Mexico seemed favourable and friendly here (similarly with a similar period Bewitched episode I watched in hospital), where have we gone wrong?

In the first season, it mostly falls under the category of competence porn, with some occasional incompetence thrown in, especially when it’s personal. (Unfortunately it’s always personal, and every member of the main cast has been kidnapped, or held hostage, or both, more than once). The problem as always in Hollywood is bad (uncreative) writing, probably because that’s what the studios want, probably because they think that’s what audiences want. Who am I to say what the US broadcast market c.1968-1974 wanted? And kidnapping and bombings in San Francisco were ripped from the headlines, if too often. Later seasons show a deterioration in competence, correlated with the deterioration in writing. To be clear, the first season is the best so far, the rest is just obsessive completionism, it’s not like modern TV with much of a story-arc, though occasionally some characters reappear, and there are a few special episodes..

FIN

Read more

Posts

Papers

• Wang, Z., Ramezani, M. and Levinson, D. (2027) A communication-free decentralised two-dimensional trajectory controller for autonomous vehicles on a lane-free road: A level-k game approach. Transportation Research Part C: Emerging Technologies, 194, 105959.

• Mann, I. and Levinson, D. M. (2026) Fare enough: A quantitative spatial transport model for transit pricing. Transportation Research Part A: Policy and Practice, 214, 105237. [doi]

Books

• Somwrita Sarkar, Unequal Agglomerations: Urban Structure, City Size, and Spatial Justice (Palgrave Macmillan). City size, inequality, and accessibility, examined through complex systems and network science.

• Jiangping Zhou, Big and Open Data for Transit Planning (Routledge). Rethinking transit catchment areas using smartcard data, OpenStreetMap, and other big and open data.

Pointers

• Your regular reminder: Car into multiple people in Newcastle NSW

• Spokane launches Emergency Streets pilot to combat fatal crashes. ALSO Emergency Streets Pilot Program.

The Emergency Streets Pilot Program is a new initiative that brings temporary traffic-calming measures to the site of a fatal traffic crash within 48–72 hours. Residents and visitors will see highly visible treatments such as cones, speed feedback signs, roadway stenciling, and bright pink signage. The program encourages drivers to slow down while the City monitors traffic speeds and gathers community feedback about the program and traffic safety. These temporary installations (deployments) will remain in place for approximately two weeks as part of the pilot.

Emergency Streets is built on the belief that traffic fatalities are preventable and that immediate action can help reduce the risk of future crashes. The pilot aims to create safer streets, build awareness about the role of speed in traffic safety, and inform future roadway safety improvements across Spokane. Fatal traffic crashes will be evaluated by the Traffic Fatality Review team, a multidisciplinary interagency group led by the City of Spokane Division of Transportation and Sustainability.

• India’s Homegrown EV Revolution (Gift Link)

• Bloccati nel traffico, cambiare corsia conviene? | Fleet Magazine1

Why Changing Lanes Is of Little Use

The same principle helps us understand what happens when, in an already congested situation, we continually move from one lane to another. Each lane change alters the space between vehicles: the car that enters can prompt oncoming traffic to slow down to re-establish an adequate distance. With light traffic, the effect is easily absorbed; near congestion, however, even a modest braking effort can spread to the following vehicles. This doesn’t mean that changing lanes is always useless. It’s necessary to overtake, take an exit, merge, or negotiate a narrowing. It’s a different matter to do it repeatedly simply because the next lane appears to be moving faster at that moment.

Yang Gao and David Levinson of the University of Sydney measured this phenomenon in real-world traffic. In a study published in 2023 in Communications in Transportation Research, the two researchers analyzed vehicle trajectories on Sydney’s M1 motorway.

Under most conditions observed, lane changes did not produce a significant speed advantage. However, a benefit did emerge in the 45 to 50 km/h range. The study also found a reciprocal relationship between congestion and the frequency of lane changes: heavy traffic pushes drivers to change lanes more frequently, and increased lane-changing activity can, in turn, contribute to worsening traffic flow. The difference between two lanes, after all, can last only a few seconds. One lane moves forward, then slows; the adjacent lane catches up shortly thereafter. Those who continue to move therefore risk chasing a temporary advantage.

• Amusements

  • The Etymology of Cars [Video]

  • Walking Men Worldwide: pedestrian traffic signals from around the world.

  • Draw all the streets in a city [via Simon Kuestenmacher].

  • LimeBike Sculpture, Sydney. I assure you I didn’t do this. LimeBikes are heavy.

• Financing

  • Private equity-backed Florida railway company Brightline files for bankruptcy.

• AVs

  • Aftermarket driver assist under federal probe following fatal crashes.

• Academia

  • Seeing Like a Gardener: A New Framework for the Public Funding of Science.

  • Could AI Be Better at Peer Review Than Humans? [Yes, no doubt] (free article, but registration required).

Press

• New Bus Routes, More Rail Services Dominate Inner West Council Transport Plan [report]

  More detail needed – expert David Levinson is Professor of Transport at the University of Sydney School of Civil Engineering. He noted that “all of these things are fine to want, but it is not very visionary,” and the council’s vision lacks detail.

  “They punted on asking for light rail on Parramatta Road, rather than say, envisioning buses now, including bus lanes and traffic signal priority for buses, and light rail later. Same with Victoria Road, they should have called for light rail. It is not like they are paying for it, but they seem to be conceding the case before asking for what they want.”

  “They missed an opportunity to call for lower speed limits on streets in residential and commercial areas and around schools, and other traffic-calming and pedestrian improvements to make it easier and safer to walk around for transport, not just recreation.”

  “Buses on WestConnex is good, and it was promised by the proponents of WestConnex when it was approved. Councils should not have to ask for things that were already committed.”

‘Crushed’: Push to punish hoons with loud cars

David Levenson [sic], a professor of transport engineering at the University of Sydney, said loud vehicles were “a serious issue” that drive down property prices.

“People who have houses on noisy streets don’t find that they can sell those houses for as much as people who have house on otherwise quiet streets.”

“And so, every unit increase in decibels results in about a 0.62 per cent decrease in the price of a house.”

In recent years, New South Wales has recorded the following statistics:

Homicides:

There were 85 murders in 2024, the highest annual count since 2014. In the 12 months leading up to March 2026, there were 87 recorded murder victims. (See: On Bondi Beach)

Road Deaths:

• Road Fatalities: In 2025, 355 people died on NSW roads, vs 327 in 2024. Speeding was associated with 134 deaths in 2025.

• Pedestrians: 54 deaths, representing a 25.6% increase from the previous year. The majority of the 54 pedestrian deaths occurred on everyday suburban streets. Older road users (aged 70+) were particularly affected, accounting for 18 of these pedestrian fatalities.

• Cyclists: 15 deaths (including three e-bike riders), which was a 200% increase over the 5 fatalities recorded in 2024. Cycling has notably increased with the advent of e-bikes in recent years.

These increases occurred despite (because of?) broader national efforts to improve safety for vulnerable road users.

The state government simultaneously blames drivers for driving too fast while opposing safe speed policies, and opposing the use of regulatory devices like unmarked speed cameras.

FIN

Read more

For my international readers, while noise is an issue in many places, in Australia this is an issue because of the popularity of hooning, anti-social operation of vehicles, especially noisy driving (modified engines, tyre burnouts (spinning the driven wheels while holding the car still), etc.) Wikipedia reports:

A hoon (/huːn/ ^ⓘ) is an Australian and New Zealand term describing a person who deliberately drives a vehicle in a reckless or dangerous manner, generally in order to provoke a reaction from onlookers.

Decibels

If a tree falls in the forest and there is no-one there to hear it, does it make a sound? Of course.

But does it make a noise? No.

Noise is unwanted sound. It is a negative externality; it requires both someone making the noise and someone hearing it without wanting to.

What matters to the person trying to sleep is how loud each event is and how many events there are.

In contrast with intuition, decibels are logarithmic. One truck may be 80 dB, but two trucks are “only” 83 dB. Ten decibels more sounds about twice as loud, so a modified motorcycle at 90 dBA is heard as sixteen times louder than a 50 dBA street.

A quiet street with one loud vehicle is worse than a busy one with a steady stream of vehicles. The background is low and every pass stands out.

Money

We have heard from my colleague Edmund Lau at the University of Sydney about the health effects of poor sleep, resulting from noise.

But everything in Australia is about real estate. Hedonic studies of real estate prices near highways and airports give us a Noise Depreciation Index. Each additional decibel lowers the price of a house by about 0.62%. People don’t want to live near highways or under flight paths, and require a lower price to do so.

Roads that invite it

Some roads invite anti-social driving: long straight sections with clear sightlines, capacity to spare at night, connected loops that let a driver circulate and repeat the performance. Hard reflective surfaces, façades close to the carriageway, and light background traffic amplify the noise.

Re-design

Design works every hour on every driver. Geometry that enforces itself changes what speed feels natural, and it narrows the whole speed distribution, including the fast tail that does the damage. Design changes may comprise various sorts of traffic calming (but not all of them), including interventions like planting trees on the side of the road, to slow traffic.

Isolated humps and stop signs produce the noisiest pattern available. Brake, sprint to the next device, brake again. The braking, the suspension and the re-acceleration all make noise.

Continuous low-speed geometry is quieter: close-spaced plateaus, raised intersections, and small roundabouts, along with traffic signals timed to give a green wave at the target speed rather than a clear run. Diverters cut flows and speeds together.

Lower speeds also lower the level. Zurich dropped limits from 50 to 30 km/h and measured about 1.6 to 1.7 dB less at building façades, with less annoyance and less disturbed sleep reported.

Re-surface

Transport for NSW notes that against dense-graded asphalt: new open-graded asphalt −2 dBA, stone mastic asphalt −1, 7 mm chip seal +2, 14 mm chip seal +4. Road pavement gets resurfaced on a cycle, we might as well put down quieter pavements when we can.

Enforce

Enforcement is our go-to policy, and we cannot abandon it. It is necessary for modified exhausts, organised racing and repeat offenders, and cameras and patrols can follow the problem as it moves. But we must also recognise it is selective by nature. And it addresses the problem that previous bad designs have created, and hopefully future good designs will prevent.

What Parliament can do

Three things, none of which requires another study.

1. Redesign. Self-enforcing 30 km/h design at noise hotspots, and on new residential, school and shopping streets. Fund calming along whole corridors, not single blocks.

2. Resurface. Fund low-noise surfaces, and specify them where people live.

3. Enforce. Objective noise limits that can be measured at the roadside, backed by cameras, patrols and inspections aimed at modified exhausts, organised racing and repeat offenders.

All of which is to say: enforcement catches the offender, design removes the opportunity.

FIN

Notes

Electric vehicles

Electric propulsion is quieter, and below about 30 km/h that is most of the sound. Above it, tyre-road noise dominates and an electric car sounds like any other.

Read more

Universities select students, enrol them in classes, and certify the result, and confuse success at one activity with success at the others.

Selecting

Do we really think firms hire Harvard graduates because of what they were taught in Cambridge rather than who was selected in the first place? Harvard has already done a screening exercise for the employer. So long as a student can comply sufficiently with the university system, they are awarded a degree four years later. That degree carries the screening information even before we ask what happened in the classroom.

Universities and their associated admission processes are designed to sort prospective students into the best institution they can attain, subject to constraints of money, geography, admissions, and everything else. Then we rank the institutions using the outcomes of the people they selected.

A university can admit excellent students, who applied because of the university’s historically accumulated reputation, and graduate excellent students without having added very much.

Another can do a great deal for students who arrived less prepared and still look worse in the rankings.

Teaching

We can document that teaching took place. Classes were taught, assignments marked, rubrics applied, student surveys completed.1 There are Canvas engagement snapshots and peer observation notes. The file is complete.

Accreditation adds its own bureaucracy. We map assessments to learning outcomes, learning outcomes to graduate attributes, and assemble reports showing that the mappings have been reviewed. External reviewers examine the evidence. We demonstrate that we have processes for assuring ourselves that students have learned.

But an accredited course is not an independently examined graduate. Reviewing our assessments and marking practices still leaves the question of what graduates retain and can do without our prompts, rubrics, and rehearsed problems. The rubric may focus the student and the instructor on more specific knowledge at the expense of the general, or other specific knowledge not previously disentangled. By over-reliance on standards, we are treating education as if it were an industrial process, rather than a bespoke set of interactions that shape individual brains.

How much of the effort spent documenting that we assess learning could go into finding out what actually survives? We have made the assurance of learning into a substantial administrative activity. We should demand evidence that the assurance itself is worth the work, we do more now than 30 years ago, but can’t show graduates are better.

What did the students learn? What do graduates still know a few years later? Can they think better? These are the actual outcomes. They are not explicitly assessed. We may know where our graduates are. We already follow graduates into employment and track their salaries. We can ask about retained learning too. A salary and an understanding of a subject are different outcomes. The rest, our assessment of teaching, is paperwork about the process.

Certifying

A passing grade records achievement at the time of assessment. Those assessments are developed by the instructors themselves. An assessment does not establish that the knowledge was retained, or that the student can use it when the problem looks different. We make a considerable leap from completing the course to having acquired an education.

A rank records relative position, not an absolute standard. Somebody has to finish last. They need not be incompetent. Somebody has to finish first. They need not be competent.

Would you cross a bridge designed by the best student in a class where nobody understood structural mechanics?

The student wants to pass. The university wants the student to remain enrolled and complete. Our officials celebrate higher passing rates. Society wants the qualification to mean something.

When those interests diverge, whom do we serve: the customer, or the society subsidising the education and relying on its standards? This arrangement deserves more scrutiny than it gets.

Assess the achievement

Society should demand the separation of teaching from at least some of the assessment.

An independent assessment should ask whether the student can do what the qualification promises. Where they learned it should be a different question. If they learned from books, work, or another teacher, the knowledge is still knowledge, the capability exists.

This would also make the teaching itself easier to evaluate. Establish what students can do when they enter. Examine what they can do when upon exit. Follow up with a sample later and find out what survived.

For example, give them a problem they have not rehearsed and require them to explain it orally. An engineer should be able to reason through a design, identify what could go wrong, and defend the assumptions. We should be interested in the student’s understanding as well as the submitted document.2

The problem is that accreditation locks in old knowledge. New material must displace the old, not just accumulate. What makes a good transport engineer in 2026 should be different from what was standard when I was a student.

But independent of the specific curriculum, the point here is to establish what the test-taker knows and can do. Comparing that with what they knew on entry gives us evidence about what the education added. We should not assume the conclusion before doing the assessment. That would not survive the most basic academic peer review.

And the lectures?

If you had asked me in the 1450s about Gutenberg’s printing press, I would have thought it sufficient to decimate traditional higher education. Once you know how to read, why do you need lectures?

Demonstrations, discussions, tutoring, laboratory experiments, sure. But lectures? The oral tradition should have fallen by the wayside centuries ago.

Will AI finally do what the printing press did not? I mean, who knows? Are universities on a long slow decline, or are we operating in the world of cartoon physics?

Read more

Thesis Title: Detecting Dangerous Driving

Lead Supervisor: Professor David Levinson.

Abstract:

Risky driving behaviors contribute to road crashes, which impose major social and economic burdens. To tackle this phenomenon, this thesis develops two main ideas. First, it examines what factors influence the choice to drive after visiting alcohol-serving establishments (ASEs), where driving may be riskier due to impairment, late-night travel, fatigue, and reduced attention. This then motivates the rest of the thesis, which uses computer vision to detect and measure dangerous and unusual driving behaviors from traffic-camera footage, linking this early-stage risk context to objective indicators of near-crashes and crash risk.

In this thesis, we begin by examining the factors influencing the choice to drive from alcohol-serving establishments (ASEs) using the Victorian Integrated Survey of Travel and Activity (VISTA) dataset. The thesis then shows how computer vision and existing roadside traffic cameras can be used to: (1) detect dangerous driving behaviors such as speeding, short headway, lane violations, and aggressive driving; (2) compare road safety indicators under different weather conditions; and (3) detect unusual driving behaviors such as slipping, running off the road, passing on the shoulder, and abrupt stopping.

The results of the first part of the thesis show that people tended to drive for longer ASEs and walk for shorter ones. Going home made driving more likely, while going to other social activities did not show the same pattern. Men were more likely than women to drive after visiting ASEs. In addition, short travel distances (<1 km) and 3–4 hour activity durations were associated with a greater chance of switching from driving to other travel modes for ASE trips.

With regard to the application of computer vision, first, data extracted from 258 hours of traffic-camera footage in Minnesota are combined with crash records from 2016–2022. The relationship between dangerous driving behavior indicators (DDBIs) and the number of events where time-to-collision (TTC) falls below two seconds (NTTC2) is examined using an Ordinary Least Squares model. A Negative Binomial Regression model then links NTTC2 to crash frequency, and Structural Equation Modeling is used to describe the broader links between driving behavior, near-crash risk, and crashes. Results show that short headway, speeding, and aggressive driving increase NTTC2, and higher NTTC2 is associated with more crashes.

Second, surrogate safety measures (modified TTC, short headway, and lane violations) are compared across weather conditions using 268 hours of traffic-camera footage collected in Minnesota. To better capture changing road-surface conditions, the road–tire friction coefficient is incorporated into the TTC and safe-gap calculations, providing a more realistic safety assessment for dry, wet, and snowy roads. The findings provide consistent evidence that adverse weather increases the risk of surrogate crashes, helping to clarify previously mixed results in the literature.

Third, this thesis develops a heuristic, trajectory-based algorithm to automatically detect unusual driving behaviors using vehicle trajectories extracted from traffic-camera video. The algorithm is applied to 360 hours of footage from nine highway segments in Minnesota under clear, rainy, and snowy conditions, covering 1,120,751 vehicles. Fifteen types of UDBs are identified, including running off the road, slipping in snow, abrupt pull-overs, and police-related stops. The results confirm that the proposed algorithm can effectively detect unusual driving behaviors from large-scale traffic-camera data.

By analyzing road traffic video footage, we demonstrated the potential of computer vision to detect dangerous and unusual driving behavior. Our findings provide a practical foundation for real-time traffic surveillance and proactive road safety monitoring.

The first publications from the Dissertation are:

• SHAER, A., FIELBAUM, A., LEVINSON, D. (2026) Detecting Dangerous Driving Via Computer Vision: Linking Video-Based Indicators To Road Crashes. Journal of Transportation Safety and Security. [doi]

• SHAER, A., FIELBAUM, A., LEVINSON, D. (2024) Choosing To Drive From Alcohol Serving Establishments. Traffic Injury Prevention. 25(8), 1013-1022. [doi]

• Wang, Z., Ramezani, M. and Levinson, D. (2027) A communication-free decentralised two-dimensional trajectory controller for autonomous vehicles on a lane-free road: A level-k game approach. Transportation Research Part C: Emerging Technologies, Volume 194, January 2027, 105959. [doi]

This paper proposes a two-dimensional trajectory controller (2DTC) for autonomous vehicles (AVs) operating on a lane-free freeway. Without relying on inter-vehicle communication, each AV independently determines its acceleration and steering in real time. The 2DTC is formulated as a receding-horizon optimal control problem (OCP) with nonlinear dynamics, costs, and constraints. The OCP is discretized and solved using the continuation/GMRES method to obtain the optimal control values. Strategic interactions between AVs are captured through a level-k game theory model. The performance of the 2DTC is evaluated through both microscopic and macroscopic analyses. Microscopic simulations show that the generated trajectories are smooth and compliant with constraints, while macroscopic measures are used to construct fundamental diagrams, demonstrating capacity levels exceeding those typically reported for human-driven freeway traffic. Further experiments show that higher reasoning depth yields reductions in travel time, while increased vehicle size heterogeneity improves traffic efficiency by enabling more compact packing in lane-free flow. In addition, more homogeneous vehicle speeds tend to result in higher average speed, thereby improving efficiency. Comparisons with a similarly defined lane-based controller reveal that lane-free traffic is less efficient when vehicle sizes are uniform but becomes more efficient once size heterogeneity exceeds a critical threshold of roughly 22%.

Keywords:

Automated vehicle, Trajectory planning, Model predictive control, Nonlinear control, Game theory

• Mann, I. and Levinson, D. M. (2026) Fare enough: A quantitative spatial transport model for transit pricing. Transportation Research Part A: Policy and Practice, Volume 214, December 2026, 105237. [doi]

Standard transport appraisal treats fare regulation in isolation, ignoring how price signals reshape urban geography. We bridge regulatory practice and quantitative spatial economics by constructing a Quantitative Spatial Transport Model (QSTM) for Greater Sydney, endogenising household and firm locations in response to transport shocks. Our framework extends the canonical quantitative spatial model by endogenising commuter mode choice and internalising network externalities through road congestion and the Mohring effect. Evaluating three public transport pricing interventions, including zero-fare, flat-fare, and fare-cap policies, we find that free transit is welfare-suboptimal, driven largely by operator revenue losses. In our setting, distance-based fares provide a spatial price signal that helps internalise resource costs and discourage sprawl, a signal that uniform flat fares mute. The model demonstrates that allocative efficiency requires a higher user contribution, revealing that weekly fare caps function as implicit subsidies for long-distance commuters. Higher fares trigger a spatial reorganisation where residents centralise to minimise commuting costs and firms decentralise to access labour. Ultimately, efficient distance-based pricing promotes a compact residential form, encourages polycentric employment, and reduces operator resource costs.

Keywords:

Quantitative spatial models, Transport appraisal, Fare regulation, Agglomeration economies, Spatial general equilibrium, Urban economics

Posts

Publications

Pointers

• Your Regular Reminder

  • Car ramming in Normandy, France

• AVs

  • Waymo involved in fatal crash, but probably not the cause

  • Waymo lifts the lid on the ‘brain’ powering its robotaxis

• EVs

  • The state of battery-swapping EVs

• Walking

  • India’s Supreme Court declared in June that the right to walk safely on designated footpaths was a fundamental right.

• Transit

  • Why transit suddenly costs so much. - by Reece Martin

• Social Media

  • Mastodon: Transport Starter Kit 1, Starter Kit 2

Press

• Plans for high-speed rail line from Sydney to Newcastle just got one step closer

Sydney University transport professor David Levinson warned against people getting excited that the plans would result in a high-speed train line being constructed.

“They still have to come up with the money to build the thing. Someone has to fund $90 billion,” he said. “Until they show the money it’s not real.”

My general practitioner (GP) was concerned about my cholesterol1 levels, particularly given my family history. He had prescribed statins. With more than 1/4 of adults in my age group on statins, this was not uncommon.

Statins did not go well. I am apparently statin-intolerant.2 They made me feel like I had a constant fever: muscle aches, tired, and what I imagine being 20 years older feels like, what the doctor called myalgia. We tried another statin to no avail.

So, rather than continuing to speculate about how much of a problem my cholesterol actually was, my GP sent me for a CT scan3 to look for calcification in the coronary arteries.

The result on that front was mild calcification on left anterior descending (calcium score of 50) (65th percentile).4 So the cholesterol level was not a major problem, all things considered.

The scan did, however, find something else.

There’s a little black spot on my heart today

The CT showed a suspected cardiac myxoma in my left atrium.

What’s a cardiac myxoma?5

It is a rare,6 usually benign tumour that grows inside the heart. “Benign” is doing a fair bit of work in that sentence. It means the tumour isn’t cancerous. It does not mean that having a large tumour floating around inside your heart is a good idea, or even not harmful.

An echocardiogram corroborated.

Mine was initially estimated at about 50 mm across and occupied a substantial fraction of my left atrium. This provides definitive evidence that I do, in fact, have a heart. It also suggests that mine is considerably more efficient than everyone else’s, given how well it had been doing its job while being starved of space.

The problem with a myxoma is not simply that it takes up space. If it gets large enough, it can obstruct blood flow through the atrial valve. Also, they are “friable,” meaning bits can break off, enter the bloodstream, and go places where bits of heart tumour really ought not go.

In retrospect, it also provided an explanation for health problems I had been experiencing that nobody had previously connected to my heart. My doctor had not been looking for a myxoma. There was no particular reason he should have been, given their rarity.

One week after it was discovered, I was in Strathfield Private Hospital having open-heart surgery to remove it.

Everybody has an open heart

I had never had major surgery before. I also don’t think I have taken a real sick day in my academic career. I have taught through pneumonia and through laryngitis, although I ended a class early that day.7

Open-heart surgery is another matter though.

Among the instructions on leaving hospital was that I would not be able to drive a car for several weeks. Ha. I also did not, however, ride a bicycle to or from the hospital.

My myxoma

The tumour was successfully removed from my left atrium. It was as expected a myxoma and benign, but after extraction measured at roughly 70 mm, confirming that removing it was the right strategy. The surgery involved stopping my heart and lungs, so I was kept alive by machines in the interim. I of course only know this because I was told, having been unconscious through the experience. Technically they opened me up twice, since there was still internal bleeding after they closed me the first time. Lovely.

Recovery from having one’s chest opened is not instantaneous. The first week in hospital was miserable, both because of the surgery and the tubes and because of being in hospital with all the machines that go ping. Walking, breathing, sleeping, coughing, and figuring out how to get out of a chair became more challenging than I remembered.

But the operation went well, the myxoma is gone, I have spent a month recovering (and will spend at least another two months they tell me, but I am fine to do anything not involving heavy lifting) and I am back at on campus facing a month of computer software needing updates which for whatever reason they need to waste my time for instead of doing when I wasn’t here for a month, with the computer on.

Causal Chain

The chain of causation here.

• My cholesterol was moderately high.

• My GP worried about it.

• I couldn’t tolerate the obvious treatment.

• That led to further investigation.

• The investigation showed that the cholesterol was not, in fact, doing very much damage.

• And in establishing that, the scan happened to find a large tumour in my heart before it could cause a stroke, obstruct blood flow, or do something else unpleasant.

• So my borderline high cholesterol may have saved my life.

My thanks to my GP for following up something that would have been easy to dismiss, and to the cardiologists, surgeons, nurses, physiotherapists, and other staff who diagnosed the problem, got me through surgery, and put me back together at Strathfield Private Hospital, and the cardio-physio group at St. George’s Hospital. Also thanks to the health care system in Australia, I am out-of-pocket well less than $AU10k for the whole thing. Also thanks to family and colleagues for picking up the slack.

3

I describe the CT scan as the worst amusement park ride ever. You lie on your back and they move you backwards and forwards while scanning you. Not painful (and not as claustrophobic as an MRI).

Read more

• Levinson, D. (2026) Testing Pedestrian Flow and Delay Estimation from First-Actuation Latency. Findings, August 20, 2026. [doi]

Traffic signal controllers timestamp the first pedestrian actuation that requests service, but they do not count the people arriving before the Walk indication. We test a theory-based estimator that treats the time from flashing-don’t-walk onset to first actuation as a censored waiting-time observation from a Poisson arrival-event process. The inferred event count and delay are calibrated separately to observed pedestrian counts and observed delay. Fourteen manually observed cases give mean adjustments of 2.202 for counts and 2.176 for delay assuming 100% compliance. A separate 24-leg-session video dataset gives 2.203 and 2.233. Held-out count errors are 23.9% and 22.4%. Observed behavior reduces delay from 16.606 to 15.996 pedestrian-hours in the manual data and from 12.934 to 8.522 hours in video.

Keywords:

pedestrian counts, pedestrian actuation, Poisson arrivals, signal delay, noncompliance, traffic signals, SCATS, Sydney

Writing and editing are, ideally, two stages in a complete creative process, yet we often prioritise the act of creation over the art of curation. The process of writing is mostly additive, with authors, starting with a blank page, spending the majority of their time on adding content rather than subtracting. This imbalance leads to verbose texts that obscure the core message, wasting the reader’s time.1.

Drawing from the metaphor of Michelangelo’s David, where the sculptor’s task was to reveal the figure already in the marble, “geological processes” of creation are akin to the writer, while the editor, our Michelangelo, must refine and shape the existing material. Sadly, Geology gets no credit despite having created the marble, and thus David.2 Yet, this editorial role is often undervalued, despite clarifying and distilling ideas for the reader, saving the reader time, lowering their costs of idea acquisition, and increasing their net value.

In the era of large language models (LLMs), which are inherently additive in nature, balancing creation with creative destruction is required to offset tools that flood the market with content quickly. LLMs are unmotivated to subtract,3 which is essential for clarity and impact. This shift from scarcity to abundance in content creation demands a renewed emphasis on curation.

FIN.

Read more

When there's an ad-wrapper on the bus you ride, 
so you can’t clearly see outside, that’s enshittification.

When someone leaves a shared e-bike 
on the footpath you hike, that’s enshittification.

When a billboard appears
right where you walked freely for years, that’s enshittification.

When the cycle runs so long 
the walkers stand till dawn, that’s enshittification.

When a car’s tuned to roar 
so the whole block can hear even more, that’s enshittification.

The 2024 word of the year, coined by Cory Doctorow,1 does not just describe internet platforms, it describes a lot of the erosion and destruction of the commons we face in urban environments, destruction at the hands of international corporations, local governments without much sense, state traffic engineers maximising vehicle throughput, or just your local neighbourhood assholes.

This is why we have to have rules and regulations and laws backed by police with guns, because too many people are selfish jerks without either any concept that what they are doing is a problem, or not caring.

Council officials who fail to do anything are complicit in the enshittification of the pedestrian realm.

All of which is to say: Operations (which deals with today’s people moving on today’s network) is more important than Planning (which deals with a small increment of additional imaginary future people moving on imaginary future networks), and we spend relatively little effort on it.

FIN

Read more