Innovative Modelling Techniques

Jan Zill, Veitch Lister Consulting - Freezing randomness at the individual-utility level in micro-simulated models

Jan Zill

Principal Consultant | Veitch Lister Consulting

As a Principal Consultant (Model Development), Jan works on all aspects of transportation model development with a focus on demand modelling. His current interest centres around the development and application of activity-based models for forecasting.

Freezing randomness at the individual-utility level in micro-simulated models

Micro-simulated econometric demand models require turning choice probabilities into a definitive outcome. In frameworks like ActivitySim, this is done by employing standard Monte Carlo methods.

However, this method is inconsistent with random utility theory at the individual level and can lead to counter intuitive results when comparing single realisations of two scenarios.

We here show how this can be avoided by freezing randomness at the individual utility level for the multinomial and nested logit models.

We implement this methodology in the open-source ActivitySim python package and use a prototype ActivitySim model of Brisbane to compare the influence of this methodology on results for a road-capacity improvement scenario.

We also investigate the influence of our methodology, which removes Monte Carlo simulation noise, on overall model convergence. This has potentially important implications for economic benefit analysis with micro-simulated models.

Lewen Feng, Monash University - An Accessibility-oriented Framework to Assess the Impact of Autonomous Vehicles on Discretionary Activity Location Choice

Lewen Feng

PhD Student | Monash University

Lewen Feng is a PhD student at the Faculty of Information Technology, Monash University. Her current research is about applying integrated transport land use modelling tools to understand the potential impacts of autonomous vehicles. Her other research interests include land use accessibility and transport demand forecasting. 

Co-Author(s):

John M. Betts

Md. Kamruzzaman

Hai L. Vu

An Accessibility-oriented Framework to Assess the Impact of Autonomous Vehicles on Discretionary Activity Location Choice

The extensive development of autonomous vehicles (AVs) is poised to revolutionise the way of travelling. Research shows that AVs might create ripple effects on various fields in land use, such as the location choice decision. In theory, accessibility is an important concept that connects transportation and land use, providing a holistic performance measure for the transport-land use system. But this concept is rarely mentioned in studies that attempt to understand the impact of AVs on location choice decisions.

To fill this knowledge gap, this paper proposes an accessibility-oriented simulation framework, called AccessSim, to study how AVs influence behavioural and land use choices. The framework consists of an activity-based travel demand model with accessibility constraints and a dynamic transport assignment model.

The accessibility constraints are situation-dependent spatial-temporal constraints derived from individuals' activity patterns. We applied AccessSim to Clayton, an inner southeast suburb in the Melbourne metropolitan area, and focused on discretionary activity location choices.

Various scenarios that involve a reduction of the value of time were tested. While most studies have concluded that AVs can significantly increase trip lengths for daily activities, our simulation results demonstrate that under space-time constraints, the increase in discretionary trip lengths is modest. Additionally, AVs can induce longer trip lengths but simultaneously increase link travel times.

Thus, the use of AVs can aggravate existing congestion and eventually reduce their net benefits. We provide a discussion of the simulation results and address the limitations of this study at the end of this work.

Peter Davidson, TransPosition - The mixed logit model - is it time to move on from logit models?

Peter Davidson

Director | TransPosition

Peter Davidson has worked in the transportation field for almost 30 years and has broad experience with transport planning and modelling with a focus on advanced analytical methods and software development.

He has developed a number of major four-step models and conceived and implemented a new approach (the TPACS model, formerly named the 4S model) which has been used for around $20b of investment decisions, including bid advice for the sale of Sydney Motorways/WestConnex, post-construction sales of Clem 7 and Queensland Motorways, business case assessment of the Toowoomba Second Range Crossing, and forecasts for Melbourne's EastLink.

Recently he has focused on understanding the impacts of autonomous vehicles, and has assisted Queensland's Transport and Main Roads in developing policies for their adoption.

The mixed logit model - is it time to move on from logit models?

Choice theory lies at the heart of any transportation model; it has been used for all stages of four step models, and for probability estimates in newer agent-based or activity models. Almost all are based on the theory of random utility maximization (RUM), first developed in the 1970's. RUM allows for a range of random utility functions, however most transport modelling is done with one of the simplest forms - the multinomial logit model. This model form does not allow for variation in taste, or correlation in the alternatives, as all of the variability is captured in a single independent and identically distributed (IID) error term.  Although the nested logit model can allow for variation between nests, the basic form of the model is very limited and arguably unrealistic. 

In the late 1990s a much more flexible model was proposed - the mixed logit model. This adopts a generalised random utility specification, which allows for taste variation and complex choice substitution. It has been widely discussed in the academic literature, and used in many research studies, but is rarely seen in any metropolitan transport models. There are a number of reasons for this, including: possible lack of understanding; increased complexity; and presumably some degree of satisfaction with the status quo.  But perhaps the biggest reason is practical - due to the complexity of the mathematics it cannot be solved analytically, and requires numerical techniques such as Monte Carlo simulation. 

This presentation will discuss the practical issues that arise when a mixed logit model is applied in practice, drawing on the author's experience with the use of mixed logit in TransPosition's Agent Cloud Simulation (TPACS) model. The practical issues associated with: parameter estimation; path-effects; convergence; and model run times are all considered. Empirical evidence of the need for better models is also discussed, as is the potential forecast biases that can result from logit models.

Mohammad Saifuzzaman, Aimsun - Aimsun Next Interface with STREAM

Mohammad Saifuzzaman

Product Specialist | Aimsun

Mohammad Saifuzzaman is a Product Specialist in the Aimsun Pty team in Sydney. He is working on product improvement, data analysis, and client support. Before joining Aimsun, Mohammad worked as an active researcher in the field of traffic flow theory and simulation.

His work has been published in top-ranked journals. With his strong research background, he participates in various research and development projects in Aimsun to improve and extend the software capability in offline and real-time traffic simulation.

Over the last twelve years, he has worked in several projects in Bangladesh, Sweden, France, and Australia.

Aimsun Next Interface with STREAM

In this project, Aimsun developed a safe and efficient digital training ground for the traffic management staff of Main Roads Western Australia to use STREAMS traffic management system. STREAMS allows operators to manage the motorway and incident response, ramp metering, and other traffic services from a single system.

Traffic managers are keen to maximise the benefits from STREAMS by training operators in every possible scenario. It is clearly inefficient to wait for incidents to happen on the road before learning how to manage them. All possible traffic management strategies need to be planned and tested in advance in a real-like environment.

Aimsun developed a novel tool to simulate the live traffic environment based on a pre-calibrated Aimsun next traffic model; this model links to STREAMS to offer a realistic test environment for the operators. Any scenario and any incident can be created in the Aimsun Next model, where operators can learn to use STREAMS to find the best possible strategies to mitigate those incidents.

The real-time traffic animation mimics the traffic monitoring cameras with some added benefits such as zooming and following the traffic. Aimsun Next replaces the physical vehicle detection system that is required for STREAMS operation with simulated detection data. It also receives any traffic management actions shown in the overhead gantries and ramp metering values in real-time from STREAMS and replicates driver behaviour accordingly.

The repeatability of the system helps the operators to re-run any traffic incident and try out different strategies. By trying different solutions within the simulation, the operator will already have a good idea of what to do when faced with a similar situation in real life. The integration between Aimsun Next and STREAMS is set up in the AWS environment, which provides easy access to both systems remotely. This concept can be used anywhere, with any piece of traffic management software.