Content uploaded by Henk Taale
Author content
All content in this area was uploaded by Henk Taale on Oct 21, 2022
Content may be subject to copyright.
EUROPEAN TRANSPORT CONFERENCE
2022
1 © AET 2022 and contributors
THE SHORT-TERM POTENTIAL OF ARTIFICIAL INTELLIGENCE
FOR TRAFFIC MANAGEMENT
Henk Taale
Rijkswaterstaat & Delft University of Technology
Erwin Walraven, Dawn Spruijtenburg, Isabel Wilmink
TNO
The field of Artificial Intelligence (AI) seems promising for traffic and transport. All kinds
of possibilities and applications are suggested, but are these suggestions feasible and
when will they become available? To address this question for traffic management, a
picture of the field and its latest, state-of-the-art innovations is painted and
opportunities for the future are investigated. Applications that have already been
implemented or tested as pilots are described, as well as those applications that
domain experts expect to be developed within one to five years, with a focus on
applications that generate the greatest improvements in terms of traffic flow, safety,
and sustainability. Also, the study looks at what the possible pitfalls and challenges
could be during development and implementation.
The research method consisted of several elements. Interviews were conducted with
experts in the field of AI and traffic management and the interviewees were asked
about possible opportunities and obstacles. In addition to the interviews, relevant and
current sources describing applications of AI in traffic management were studied. The
focus was on the added value of applications that have already been implemented.
Based on the information gathered, a selection of the most promising future
applications was made and these applications were discussed in a workshop.
The current applications of AI in traffic management show that the focus is now on
performing one specific task, using a limited number of data sources. It also shows
there is great future potential for AI-based applications that combine multiple data
sources or address multiple complex tasks in a combined fashion. This could, for
example, lead to new insights about traffic being derived from data; insights that are
not readily apparent with existing methods and a single data source.
1. INTRODUCTION
After winning the first match in 1996, the reigning chess world champion Kasparov
was beaten in 1997 in a rematch by Deep Blue II, a chess computer developed by IBM
(Wikipedia, 2022a). Although the machine falls within the classification of artificial
intelligence, it can at best be considered an expert system. Many of the parameters in
the algorithms were set by chess grandmasters, but its main strength still came from
the brute force computing power to analyse and evaluate as many positions as
EUROPEAN TRANSPORT CONFERENCE
2022
2 © AET 2022 and contributors
possible. A next step in chess computing was taken in 2017 when AlphaZero was
developed, a chess program which was given the rules of chess and was trained by
playing games against itself and learning from them (Wikipedia, 2022b). It easily beat
the best computer program at that time and was also an inspiration for many
grandmasters to try certain moves and themes which were unexplored before that
time.
Not only from the field of chess, but also from other scientific research fields we see
that artificial intelligence (AI) has developed into a main research topic and AI is
currently at the centre of scientific interest with applications being developed across a
wide range of fields. This makes it an important moment to also investigate what AI
can do for traffic management. A lot of research is currently being conducted in the
field of pattern recognition, intended to make automated driving possible. But are there
other areas within traffic management that could benefit from the possibilities AI
offers? To answer this question, a quick scan study was done to look at the potential
applications of artificial intelligence (AI) in traffic management and information in the
short term (1-5 years). This includes both current and future applications. For current
and future applications there are a number of important questions:
What applications currently exist?
What is their added value?
What opportunities will there be in the near future (1-5 years)?
What is needed to realise these opportunities?
The research was executed using various methods. Interviews were held with experts
in the field of AI and traffic management and they were asked about possible
opportunities and obstacles. In addition to the interviews, relevant and current sources
describing applications of AI in traffic management were studied. The focus was on
the added value of applications that have already been implemented. Based on the
information gathered, a selection of the most promising future applications was made
and these applications were discussed in a workshop. This paper describes the results
of the research. First, the fields of AI and traffic management are described. After that
some current and future applications for the different traffic management functions are
discussed and finally some conclusions are drawn.
2. ARTIFICIAL INTELLIGENCE
Artificial intelligence (AI) is a collective term and an interdisciplinary field in which
systems are developed that are capable of performing tasks that usually require
human intelligence (Russell & Norvig, 2020). Within the field of AI, there are several
sub-areas that focus on specific techniques. In practice, these sub-areas do not stand
alone. In many cases, applications are being developed that bring together several
sub-areas. The division into sub-areas can be done in different ways and is not
EUROPEAN TRANSPORT CONFERENCE
2022
3 © AET 2022 and contributors
uniquely defined in the literature. The following sub-areas give a global overview of
the field:
Machine learning: algorithms that are able to learn from data or from experience
(Alpaydin, 2020). This includes generating predictions or recognising patterns in
data. Supervised learning includes all methods that learn through example and for
which the desired output is known (e.g. whether or not a cat is present in a picture).
Unsupervised learning does not require this desired output and is used, among
other applications, to recognise patterns or clusters in data. Another commonly
used term is deep learning (Goodfellow et al., 2016). This is a machine learning
technology in which (deep) neural networks are trained using data.
Natural language processing: understanding and processing of natural language
(Raaijmakers, 2022). Chat bots able to understand what humans are saying are an
example of this.
Speech processing: recognising and optionally translating spoken language
(Rabiner & Schafer, 2011). This is for example used in mobile phones, able to
respond to spoken commands. Speech processing always involves spoken
language listened to by a computer, while natural language processing focuses on
the analysis of a text after a computer has converted it into data.
Computer vision: processing still and moving images, and understanding what
happens in them (Szeliski, 2010). Recognition of objects in photographs is an
example of computer vision.
Expert systems: systems that possess specific knowledge provided by human
experts to solve problems in specific areas (Gupta & Nagpal, 2020). An example is
the analysis of a disease profile using questions asked by the system and answered
by a human being.
Planning, scheduling, and controlling: systems that independently determine which
actions must be performed to achieve a goal (Kochenderfer, 2015). An example is
the control of traffic lights to maximise traffic flow. Reinforcement learning (Sutton
& Barto, 2018) also falls within this sub-area. This is a popular machine learning
technology in which a system independently decides on actions to take and learns
to function more efficiently by analysing the results of its actions. This learning
process works using a reward function, whereby the system is rewarded for
demonstrating the type of behaviour that achieves the system’s end goal.
Systems that are developed using AI are becoming increasingly complex, making it
more difficult for humans to understand how they work and why they are making
certain decisions or predictions. A current trend is the development of technologies
that make AI-based decisions explainable and more transparent (Adadi & Berrada,
2018). This should make it easier for humans to discover how an AI system has arrived
at a particular prediction or decision.
Although AI is currently receiving a lot of attention and has achieved great successes,
there are also limitations that could pose risks in its application. An example is
EUROPEAN TRANSPORT CONFERENCE
2022
4 © AET 2022 and contributors
recognising objects (e.g. traffic signs) using camera images. It appears that in some
cases it is possible to fool a data-trained AI system by making minimal (invisible to
humans) changes to these camera images. A red traffic light might, for example, be
recognised as a green traffic light, which could lead to dangerous driving behaviour on
behalf of automated vehicles using this information to make their decisions.
Vulnerabilities such as these show that the practical application of AI is challenging
and must be done carefully.
3. TRAFFIC MANAGEMENT
Although the definitions of traffic management can differ, the aim of traffic
management is to influence traffic supply and demand in such a way that traffic
demands and the capacity supply of the network are better matched, both in time and
space. The problems encountered on the road network mainly concern specific
bottlenecks and moments (i.e. peak hours, incidents, and events). By spreading traffic
demand or dynamically adapting the supply of infrastructure, the existing road network
can be better utilised. Typical traffic management measures include ramp metering,
dynamic speed limits, peak hour lanes, but also traffic information communicated
through variable message signs or other channels. The measures are primarily
intended to improve accessibility, but sometimes they are implemented also to
improve road safety (e.g. through queue tail warning) or quality of life (e.g. by using
speed limits) (Hoogendoorn et al., 2012).
Traffic management can fulfil various traffic-related functions:
Monitoring and detecting: the monitoring and detecting of traffic and incidents.
Informing: traffic signs, route information, travel times or lanes that can be used.
Advising: advise on lanes, speeds and alternative routes.
Warning: queue tail warning, dangerous situations, disturbances.
Management and control: reducing speed limits, changing lane allocations, opening
or closing lanes, processing height alerts, overtaking prohibition, metering and
buffering.
Aspects that do not fall under traffic management are mobility management (e.g.
measures to avoid peak hour traffic) and road pricing (e.g. congestion charging).
However, such measures can temporarily or locally reduce traffic flows, making them
more manageable.
The AI applications in this paper are categorised according to their traffic management
functions. Informing, advising, and warning have been combined, since these
categories all focus on transferring information to road users or traffic control centres.
In this paper not all applications are discussed, but more and different examples of AI
in traffic management can be found in Walraven et al. (2021).
EUROPEAN TRANSPORT CONFERENCE
2022
5 © AET 2022 and contributors
4. MONITORING AND DETECTING
The use of monitoring and detecting allows situations on the road to be analysed. The
information gathered can be used for decision-making or to identify connections. Think
of applications that inspect peak hour lanes (for the presence of stranded vehicles) or
monitor the number of road users at specified times. Various data sources exist (e.g.
video images) for which AI can be used to analyse large quantities of data in an
automated way.
4.1 Current applications
Detecting mobile-phone use behind the wheel using smart cameras
Reducing the handheld use of mobile phones behind the wheel is expected to increase
safety on the road. Police in the Netherlands have a number of smart cameras at their
disposal. They take pictures of passing cars and use deep learning techniques to
decide whether they suspect the car driver of holding a mobile phone. Photos on which
a car driver is suspected of holding a phone are forwarded for verification and manually
checked. When the handheld use of a mobile phone has been verified, the vehicle’s
licence-plate number is used to send the driver a traffic fine. The added value of AI in
this application is the automated and large-scale recognition of drivers on their phone.
Analysing all this data would be an incredibly labour-intensive task for humans. One
of the challenges regarding this application is that, when constructing the model,
photos need to be used that have already been manually checked. This manually
labelled data can then be used to train the model.
Inspecting peak hour lanes using smart cameras
Rijkswaterstaat also uses camera images in its traffic control centres. A lot of people
in these control centres spend a lot of time doing routine work, such as recognising
objects on a peak hour lane. This is work that people are unable to perform well for
longer periods of time (e.g. due to fatigue). To automate the recognition of objects on
the road, a system has been developed that automatically recognises objects using
video images (AI sub-area: computer vision). When developing the system, it proved
to be a challenge to test how well the system functions. Especially, whether it functions
well enough to take over tasks from people. In traffic control centres, employees are
responsible for their own decisions. When an independently operating computer
system makes a wrong decision, it can be difficult to allocate responsibility. Systems
must therefore be extensively tested before they can be put into practice.
4.2 Future application
Predicting the duration of an incident
Rijkswaterstaat has developed a system that can predict the probability of an incident.
This information is used to identify hotspots, e.g. locations where the risk of incidents
is higher than at other locations. This information is used to position road inspectors
EUROPEAN TRANSPORT CONFERENCE
2022
6 © AET 2022 and contributors
in such a way that it increases their likelihood of being in the vicinity of a potential
incident. The next relevant step is to look at predicting the duration of incidents. In
addition, it may be possible to predict how the effects of incidents spread across time
and space. Doing so can create insight into where people are likely to experience a
disturbance, and also, where measures or route advice could be useful. Linear
regression models can be used to create a prediction model. More advanced
technologies, such as decision trees and neural networks, can also be used.
Training a model on incidents is a challenge since incidents are infrequent and arise
from specific situations. Incidents in historical data can differ greatly from each other.
Since such a model is trained on historical data, it is important for it to be able to
indicate when it does not recognise a situation. Some types of incidents will be so
unusual and without precedent in the historical data, that the model is unable to
accurately make a prediction. The model must therefore be aware of its own
limitations, so that in such cases the estimate can be made by an expert. A possible
challenge when developing the model could also be that the information on incidents
is incomplete in the data. There has to be enough data on enough incidents to train
the model properly. In addition, it can be difficult to link contextual information to
incidents, while such contextual information can have a significant impact on traffic
flow during incidents. In general, AI is good at predicting large intensity patterns, such
as traffic jams. However, it is more difficult to be very accurate and predicting incidents
and their duration can be a challenge.
5. INFORMING, ADVISING AND WARNING
Informing, advising, and warning concerns information transfer to road users or traffic
control centres. The communication can be informative, but can also be in the form of
advice or a warning. Applications include route information, speed advice, and queue
tail warning. AI offers opportunities to better inform, advise, and warn road users using
multiple data sources. Combining multiple data sources allows for more accurate
predictions, from which more reliable information can be derived and presented to road
users.
5.1 Current applications
Route advice to avoid and prevent traffic jams
In the European project SOCRATES2.0, traffic predictions were used to advise better
routes to road users (Huisken, et al., 2020). The project developed systems that
provided proactive advice. Their aim was not only to avoid existing traffic jams, but
also to prevent expected traffic jams. Real-time data was combined with historical data
to calculate the probability of congestion at a specific location in the near future. If the
probability was high, alternative routes were suggested to road users. Machine
learning was used for the predictions. This system was tested in the Amsterdam region
EUROPEAN TRANSPORT CONFERENCE
2022
7 © AET 2022 and contributors
between December 2019 and the summer of 2020. However, there were fewer traffic
jams than usual during the test due to the outbreak of COVID-19, meaning reliable
conclusions regarding the effects on traffic have yet to be drawn.
Informing travellers about the opening of a bridge
The Botlek Bridge and Spijkenisser Bridge regularly open for shipping on the Oude
Maas river, resulting in longer delays on the road. In collaboration with Technolution,
Rijkswaterstaat developed a system to predict the opening of these bridges, so that
road users are better informed and the use of alternative routes can be stimulated.
The system uses information transmitted by ships to predict whether one of these
bridges will open in the near future. When the probability of opening increases, the
message ‘no opening expected’ on dynamic route-information panels is replaced by
‘opening expected’. In this way, road users are proactively informed about bridge
openings, and they can decide to take an alternative route or not. The system was
tested in practice for six months at the beginning of 2020. Regarding the predictions,
the message ‘no opening expected’ on the information panels proved to be more
reliable than the message ‘opening expected’. This is due to the fact that the exact
routes of ships on the Nieuwe Waterweg are difficult to predict.
Figure 1: Botlek Bridge and Spijkenisser Bridge (source: Wikipedia)
5.2 Future applications
Predicting travel times and advising routes
Graph neural networks are relatively new models that use a graph structure as input
for a neural network. While training, the neural network learns what the dependencies
are between the locations in the graph. It is interesting to see what potential these
kinds of models have within the context of traffic management. For example, they
make it possible to make network-wide predictions since dependencies between
locations can be taken into account more easily. In addition, contextual information,
such as the presence of events, can be included. Better predictions of network
conditions are especially beneficial in the case of incidents.
EUROPEAN TRANSPORT CONFERENCE
2022
8 © AET 2022 and contributors
The improved network-wide monitoring can be used for various purposes. Firstly, it
can serve as a monitoring tool for traffic control centres. They can be alerted to
situations that deviate from normal. If needed, they can then apply measures in the
network to improve such situations. The added value of the model is that it can be
constantly monitored and updated. In addition, predictions from (graph) neural
networks can be used as input for generating route recommendations to individual
road users. This could include advice that is in the public interest (e.g. suggesting
routes that circumvent public schools) or advice that aims to avoid vulnerable points
in the network (e.g. bridges) at times when the network is already crowded. Herein,
finding the balance between what is beneficial to individual travellers and the collective
is challenging.
When using graph neural networks, the structure of a graph can be directly taken into
account during learning. This allows for a contextual representation of the data to be
made. For this, it is important to think carefully about the way the data is projected
onto the graph. There are, for example, several ways to divide a road network into
segments. The way in which this is done affects the functioning of the graph neural
network. The scalability of graph neural networks to larger traffic networks is an open
challenge.
Shockwave damping
There are algorithms, such as SPECIALIST (Hegyi, et al., 2008) and COSCAL
(Mahajan, et al., 2015), that dampen shockwaves using traffic flow theory. They set
speed limits that constrain the flow of traffic upstream, thus homogenising traffic speed
and dampen shockwaves. Sometimes additional measures are needed upstream,
because this is the location where higher levels of traffic density arise. AI could be
used to create algorithms that resolve shockwaves and also reduce their unwanted
effects. One possible AI technology is reinforcement learning, which can learn how to
‘squeeze’ the traffic most effectively. A reward function in reinforcement learning could
be expressed in vehicle hours of delay. In this way, the system can learn to avoid
congestion.
It is expected that AI can contribute to the improvement of algorithms in a number of
ways. First, it may be possible to develop better decision rules using reinforcement
learning. Rules for the speed of advice could improve, but also the activation of the
system. In many cases the system is not activated, even though there are
shockwaves. The conditions under which the system activates are strict and presently
often remain unmet. When correcting this, care should be taken not to set the threshold
of activation too low. If the system kicks in at every minor traffic flow disruption, it loses
credibility. In addition, the type of intervention is currently determined only once, upon
activation. Doing so more dynamically, while adding the option of making adjustments
during the intervention based on the reaction of the traffic, could enhance the system’s
performance.
EUROPEAN TRANSPORT CONFERENCE
2022
9 © AET 2022 and contributors
However, it is important that the self-learning system does not start without knowledge.
The current algorithms can be implemented as the basic knowledge that the system
can improve upon. Certain safety constraints must also be imposed on the algorithm
to prevent the system from engaging in undesirable behaviour (such as constantly
changing speed limits). New developments in the field of safe reinforcement learning
can offer a solution here. Before such a (renewed) system can be used on the road, it
is necessary to show what its added value is in different traffic situations (also
compared to existing systems). This can be done in simulations.
6. Management and Control
Management and control concerns intervening in the way traffic is guided, to optimise
the flow of traffic. Applications include reducing the speed limit or opening and closing
a lane. Management and control can also be done by models. AI methods can play a
role in such applications, especially in making decisions to redirect traffic without direct
human intervention.
6.1 Current applications
Adaptive traffic lights
AI is also used to control with the help of the existing infrastructure. AI algorithms are
used, for example, to determine when traffic lights turn green and for how long they
remain green in areas with multiple intersections. SURTRAC is a system that plans
the control of traffic lights for a specific intersection based on information regarding
approaching vehicles (Smith, et al., 2013). The information from an intersection is then
communicated to intersections in the vicinity, so that it can be used during planning.
SURTRAC uses AI-based planning and scheduling. It can be seen as an ‘online’
controller that constantly optimises itself as new information comes in. SURTRAC is a
multi-agent system, in which multiple autonomous agents exhibit smart behaviour and
cooperate. This is a field at the intersection between artificial intelligence, optimisation,
and operations research. SURTRAC was tested at nine intersections in Pittsburgh
(Pennsylvania, USA) and compared to a reference controller that does not
continuously optimise. The results show that travel times can be reduced by an
average of 26 percent. Emissions were also reduced by 21 percent during the test.
Additionally, road users were able to drive faster and had to stop less often.
Determining the positions of road inspectors
Road inspectors are deployed when incidents are reported. They are responsible for
ensuring traffic safety and maintaining a certain level of throughput during an incident.
Incidents may concern accidents but could also include, for example, a stone on an
emergency lane or a broken down vehicle. Therefore, it is important for road inspectors
to be able to arrive at the scene quickly. Rijkswaterstaat recently started using a new
system to optimise the position of road inspectors using incident predictions, with the
EUROPEAN TRANSPORT CONFERENCE
2022
10 © AET 2022 and contributors
aim of reducing their incident response time. This process consists of two steps. In the
first step, a Bayesian model is used that predicts the probability of an incident. This
model is estimated using data on historical incidents from a whole year. The
predictions make it possible to identify ‘hotspots’ (areas where the probability of
incidents is high compared to other locations). In the second step, the predictions are
used to optimise the positions of road inspectors. It is important for road inspectors to
be in the vicinity of locations where a relatively high probability of incidents is predicted.
Finally, this information is passed on to the road inspectors, allowing them to
reposition.
The system is currently active in the central region of the Netherlands and will soon
be deployed in the south/west of the Netherlands as well. Before the system was
implemented, the average response time was 18 minutes (measured over 24 hours);
after the system’s roll-out, it was 14 minutes.
6.2 Future applications
Self-learning traffic signal controllers and control systems
Reinforcement-learning methods are capable of making computer systems learn how
to perform a task in the best way. They can also be used to enable already installed
systems to adapt to new situations. The expectation is that continuously learning traffic
signal control or ramp metering systems will control traffic better than the traditional
systems (i.e. not self-learning) systems. This could result in better traffic flow at
intersections. The self-learning aspect allows for continuous optimisation. For
example, the construction of new residential or business areas could lead to new
traffic. Traditional systems have to be tuned manually to the new situation and this
takes time and research. Using the self-learning function of the traffic signal controller,
the signal plans are adapted automatically to perform the best for the new situation.
Traffic light controllers can be trained using historical data in combination with
simulation-generated data. Reinforcement learning operates by evaluating random
actions. As such, it cannot be directly tested in practice. The development of safe
reinforcement techniques, in which random actions are defined within certain
frameworks, could change this in the future.
Predicting the effectiveness of an overtaking prohibition for trucks
Prohibiting trucks from overtaking means passenger cars are forced to slow down less
often. This can make the traffic more stable and less likely to become congested.
Cameras make it possible to monitor which vehicles are driving on which lane and
what the traffic situation is at that moment. In addition, it is possible to monitor the past
effectiveness of an overtaking prohibition for trucks. By collecting a lot of data, it is
possible to train a model to predict for a given situation whether an overtaking
prohibition will have a positive effect on traffic flow. This model can then be used to
decide whether or not to impose an overtaking prohibition for trucks.
EUROPEAN TRANSPORT CONFERENCE
2022
11 © AET 2022 and contributors
The expectation is that a model can be constructed based on (convolutional) neural
networks, which, using images and other information, can assess whether overtaking
prohibitions can have a positive effect. When training the model, it is necessary to
distinguish between regular congestion and disruptions caused by overtaking trucks.
From this, the model must distinguish whether a overtaking prohibition is effective.
How this is to be derived from the data is still an open challenge. Aside from the
distinction between regular congestion and congestion caused by overtaking trucks, it
is also important to investigate which types of data contain sufficient information on
vehicle types and vehicle positions per lane. Therefore, as a first step, it is important
to investigate whether a connection can be made between overtaking trucks and
congestion.
7. Conclusions and recommendations
Artificial intelligence is currently in the centre of attention and is seen as a collection
of technologies that enable new applications in various domains, in which people are
supported or processes are fully automated. In traffic monitoring, new developments
in AI allow more data to be derived from existing and new data sources, such as
camera images. In recent years, there have been many developments in convolutional
neural networks that can learn from images. This will soon enable new monitoring
applications that were previously difficult to achieve, such as automatically recognising
objects in images. In addition, there are future opportunities for making more network-
wide predictions. Moreover, developments in federated learning have the potential to
provide insights derived from privacy-sensitive or business-sensitive mobility data
originating from multiple parties.
Regarding informing, advising, and warning, AI offers opportunities to proactively
present information to road users. Applications in many cases consist of two phases:
in the first phase, AI is used to predict; in the second phase, these predictions are
used to decide which information to present to road users (e.g. the advice to take
another route). It can be relevant for road users to know why specific advice is being
given (e.g. to increase compliance rates). It is expected that new technologies in the
field of explainable AI will contribute to this in the coming years.
Concerning controlling traffic, AI makes it possible to develop systems that can
independently decide how to control traffic (e.g. to improve traffic flow). Current traffic
management systems usually cannot adapt automatically to new situations. AI-based
systems can be made to be adaptive and continuously learn (e.g. through
reinforcement learning). This also offers advantages in the design and development
of these systems. For example, it used to be necessary to determine all the decision
rules for controlling traffic in advance. AI will make it possible in the future for systems
to independently learn to control traffic by making decisions and observing the results
EUROPEAN TRANSPORT CONFERENCE
2022
12 © AET 2022 and contributors
of these decisions. These self-learning systems can be used for traffic control, ramp
metering, or to decide whether to impose overtaking prohibitions.
The current applications of AI in traffic management show that the focus is now on
performing one specific task using a limited number of data sources. This makes
sense, given the fact that such applications are usually developed incrementally.
There is great future potential for AI-based applications that combine multiple data
sources or address multiple complex tasks in a combined fashion. This could, for
example, lead to new insights about traffic being derived from data; insights that are
not readily apparent with existing methods and a single data source. However, to make
AI applications successful, it is needed to include domain knowledge in the
development of models and algorithms. AI is capable of deriving insights from data on
its own, but integrating domain knowledge will always be necessary, even for the
simplest predictions.
The development of AI-based methods starts with high-quality data, making it of great
importance that the data that enables models to make accurate predictions is
available. Currently, there is a strong focus on the development of AI technologies.
However, it is important to also focus on the collection, pre-processing, and fusion of
detailed traffic data, which can then serve as the fuel for models. AI models that are
constructed using data can draw conclusions about situations similar to those seen in
the data, but at present, they are not advanced enough to draw conclusions about
situations not previously observed in the data. Combined with data, AI can make a
great contribution to traffic management. It is important, however, that AI is not seen
as a technology that always has the solution, especially regarding situations about
which little information is known.
New developments in AI in recent years have made models and algorithms
increasingly complex, which has made it harder for domain experts to understand AI-
based decisions. It is crucial for models to be explainable. It is important to emphasise
that this explainability has limits, especially when complex models are used to make
accurate predictions or decisions. Applications that demand a high degree of
explainability with a lot of detail, may be better served by models that are less complex
but more explainable. A downside of this is that simpler models may reduce the quality
of predictions. There will always be a trade-off between the quality of predictions and
the degree of explainability. On top of explainability, there is increasing attention to
biases, which can lead to undesirable decisions in practice, in data and models. It is,
for example, undesirable for specific groups of road users to never get a green light at
an intersection.
All-in-all, there is much potential to apply AI in traffic management. Therefore, it is
valuable for experts in traffic management to learn more about the possibilities offered
by AI. The expectation is that it will be crucial for traffic management experts to actively
engage with those in the field of AI. In this way, new developments can find their place
EUROPEAN TRANSPORT CONFERENCE
2022
13 © AET 2022 and contributors
in the domain, and we can be sure that new developments in the field of AI are driven
by the questions arising in practice.
BIBLIOGRAPHY
Adadi, A. and M. Berrada, Peeking Inside the Black-Box: A Survey on Explainable
Artificial Intelligence (XAI), in IEEE Access, vol. 6, pp. 52138-52160, 2018, doi:
10.1109/ACCESS.2018.2870052.
Alpaydin, E., Introduction to Machine Learning, The MIT Press, March 2020.
Goodfellow, I., Y. Bengio and A. Courville, Deep learning, The MIT press, November
2016.
Gupta, I. and G. Nagpal, Artificial Intelligence and Expert Systems, Mercury Learning
& Information, April 2020.
Hegyi, A., S.P. Hoogendoorn, M. Schreuder, H. Stoelhorst and F. Viti, SPECIALIST:
A dynamic speed limit control algorithm based on shock wave theory, proceeding of
the 11th International IEEE Conference on Intelligent Transportation Systems,
October 2008, pp. 827-832.
Hoogendoorn, S., H. Taale, I. Wilmink, R. van Katwijk, B. Immers and H. Schuurman,
The Future of Traffic Management - State of the Art, Current Trends and Perspectives
for the Future, TrafficQuest, 2012.
Huisken, G., M. Pepikj, I. Yperman, A. Feitsma, N. Rodrigues and T. Vlemmings,
Optimising Network Traffic Flow with Cooperative Traffic Management in the
Amsterdam Region, Virtual ITS European Congress, 9-10 November 2020.
Kochenderfer, M. J., Decision making under uncertainty: theory and application, The
MIT Press, 2015.
Mahajan, N., A. Hegyi, G.S. van de Weg and S.P. Hoogendoorn, Integrated Variable
Speed Limit and Ramp Metering Control Against Jam Waves - A COSCAL v2 Based
Approach, Proceedings of the IEEE 18th International Conference on Intelligent
Transportation Systems, November 2015, pp. 1156-1162.
Raaijmakers, S. A., Deep Learning for Natural Language Processing, Manning
Publications, 2022.
Rabiner, L. and R. Schafer, R., Theory and applications of digital speech processing,
Prentice Hall Press, 2011.
EUROPEAN TRANSPORT CONFERENCE
2022
14 © AET 2022 and contributors
Russell, S. and P. Norvig, Artificial Intelligence: A Modern Approach, Pearson, 4th
edition, 2020.
Smith, S., G. Barlow, X.-F. Xie and Z. Rubinstein, Smart Urban Signal Networks: Initial
Application of the SURTRAC Adaptive Traffic Signal Control System. Proceedings of
the International Conference on Automated Planning and Scheduling, Volume 23(1),
pp. 434-442.
Sutton, R. S. and A.G. Barto, Reinforcement learning: An introduction, The MIT Press,
2018.
Szeliski, R., Computer vision: algorithms and applications, Springer Science &
Business Media, London, 2010.
Walraven, E., D. Spruijtenburg, I. Wilmink and M. Schreuder, Artificial Intelligence and
Traffic Management - Current and Future Applications, TrafficQuest report, December
2021.
Wikipedia, Deep Blue versus Garry Kasparov, available through
https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov, accessed on
August 15th, 2022.
Wikipedia, AlphaZero, available through https://en.wikipedia.org/wiki/AlphaZero,
accessed on August 15th, 2022.
