With individual passenger context data available, the opportunity to truly understand an individual’s end-to-end journey experience in real-time can identify and guide tailored individual experience improvements, also in real-time. This feat is now possible by assembling available Soft AI components into viable AI solutions. Human and automated engagement can then be planned and delivered in real-time, whether services were disrupted or not.Achieving this really would be putting passengers first and making every second count.
There is a clear objective in the UK rail industry of ‘Putting Passengers First’. The challenge is multi-faceted and nuanced and includes:
- Making investments in stations, rolling stock and infrastructure that extracts the maximum value out of every pound spent;
- Establishing a regional and route centric focus that increases the understanding of passenger needs in those areas;
- Making ‘Every Second Count’ by avoiding delays, avoiding disruption and keeping services moving.
Meeting the challenge will not be easy and the potential long-term changes in society arising from the Covid-19 pandemic will change the way passengers use public transport. Yesterday’s ‘normal’ is unlikely to return and changes in travel patterns and volumes as we emerge from the pandemic will affect investment decisions. Nonetheless, deploying technology that supports ‘Putting Passengers First’ will remain a priority as will the requirement to improve service performance by making every second count.
Artificial Intelligence (AI) will play a role, not least because there may be no alternative solutions to address the most complex passenger-centric problems in real-time. The Rail Safety and Standards Board has recognised this with its Data Sandbox programme, providing access to large sample datasets sourced from across the whole rail system to look at, for example, how machine learning, graph theory and data analytics techniques can be used to make accurate predictions about the impact of reactionary delays and dwell time variations.
However, to understand the full range of applications of AI it is necessary to understand what AI and Deep Learning have to offer.
Soft AI Systems
The concepts and theory underpinning Soft AI have been around for decades. For example, the original concepts of artificial neurons and neural networks were developed in 1943 then expanded upon in the 1950s. However, the only reason AI applications are now becoming so prevalent is the low cost and wide availability of computing power, vast memory and high-speed data communication systems that can be deployed by both local and cloud-based solutions.
Soft AI applications include the type of processing, logic and interaction undertaken by Alexa, fingerprint and facial recognition systems, or playing games such as chess and GO. Programmed responses are built and configured to convert encoded input data from audio patterns, images or sensor measurements into outputs such as speech, identities or game moves. In reality, these AI solutions are just an assembly of components such as encoded data, logic, pattern recognition algorithms, control and communications systems.
Soft AI solutions are not intelligent in the truest sense, but they can deal with complexity, and with increased computing power, deal with it fast. As of today, they cannot understand the impact of the information on an individual’s experience nor why they have been built. Understanding why a question needs to be asked and how to improve an individual’s experience given their personal context at any moment in time is usually classified as a Hard AI problem requiring top-down reasoning and resembling the ways humans approach problems and tasks. Whilst designing and deploying these Hard AI systems may well be some way off in the future, in the meantime Soft AI solutions can be assembled using Soft AI components to complement human activity and improve an individual’s passenger experience
Deep Learning is often misunderstood as Hard AI. In reality, Deep Learning is another form of Soft AI where pattern recognition or forecasting algorithms that map input to output data are developed and encoded in ‘black-boxes’ such as neural networks.
The algorithm parameters that translate input data into output data are fine-tuned by iteration and embedded within a neural network; but the black-box translation itself is opaque. Whilst the system may look like it shows human-like pattern recognition or decision-making behaviour, the black-box algorithm is just a programmed response. Still, deep learning is capable of dealing with complexity where relationships between service provision and passenger satisfaction are difficult to define through rules. But care needs to be taken since if the black box has been built using biased data, its programmed response to improve passenger satisfaction will also be biased. The system does not really know what is best for passengers.
Similarly, self-organising AI systems are soft. The algorithms used to self-organise are programmed in advance, often extracting key data features in a similar way to traditional statistical techniques. This reduces ‘big data’ into much smaller sets of features. However, the final deployment of a self-organising system is still based on a programmed response.
For all Soft AI systems fine-tuning the programmed response updates can also be scheduled over time as new data or AI components become available. This is often referred to as learning, with the result being a new programmed response.
Putting Passengers First – Potential AI Solutions
Soft AI applications and systems are, in effect, an assembly of data acquisition systems, data pre-processing algorithms, hard logic, pattern recognition systems, control and communication components. These components can be assembled to process data, search for patterns and then, once configured, respond in the programmed way.
So, when it comes to Putting Passengers First, Soft AI is well suited to solving problems associated with planning, forecasting and prediction. However, in an operational environment the greatest short-term impact will come from humans and machines working together. Humans and Soft AI systems have complementary strengths: the curiosity, creativity, and emotional intelligence of humans complements the computational speed and data processing capabilities of Soft AI systems. What comes naturally to people is difficult for machines, and what is straightforward for machines, such as processing and analysing gigabytes of data in real-time, is impossible for humans.
The best AI solutions will be based on the right mix of people, processes and AI technology.
Resilient Timetable Planning
Timetables may have flaws or make sub-optimal use of rail network infrastructure. In the worst case, the planned timetable may be impossible to deliver, or is close to a tipping-point of significant disruption arising from minor perturbations. Traditional timetable plans that are valid on paper may not address resiliency problems in advance and as a result lead to unnecessary disruption that could be avoided through improved planning. The combination of sub-optimal plans with potential tipping-points reduces service resiliency.
To improve timetable resilience, historic data patterns can be collated and encoded to look at what actually happened on the rail network in the past, not just when incidents occurred but also when services ran to the scheduled timetable with minor perturbations. Combining these data patterns with the human or automatic actions taken at the time to manage performance, all of which can also be encoded, allows the AI system responses to be programmed. The human and automatic interventions that led to the best outcomes in near-identical situations can be identified and this approach can be applied at junction, station, terminal, route or even area level.
Reducing unnecessary delay by increasing timetable plan resilience is needed to make ‘every second count’. The collaboration between humans and a Soft AI solution could help build a timetable that would prevent the most common tipping-points and reduce the impact of disruption. It could also schedule recommended human and automatic interventions for different scenarios, which are hard to encode safely, with this AI augmentation improving network performance. Also, the scheduling of human and automatic interventions would be more consistent in similar scenarios, irrespective of who is involved, so the outcomes and service impact would be more predictable.
Asset Management and Failure Predictions
The failure of assets on the railway and the time taken to provide fixes is hugely disruptive to passengers. However, capturing the state of an asset and its likelihood of failure can be a low cost and straightforward exercise. Asset usage frequency, modes of usage, location, age and images are all sources of data that can be encoded to inform the assembly of a Soft AI solution for asset condition analysis. This solution can then be fine-tuned over time as more asset and failure data becomes available.
Once the data is encoded for assets such as points, signals, tracks and structures; different types of asset failure modes can be recorded and the patterns of data leading up to the failure analysed. Once the time-series data in advance of the failure and the time of the actual or likely failures are known, the input patterns leading up to a failure can be identified. This provides a near real-time measure of asset condition and the likelihood of different failure modes occurring as the asset ages. Preventative maintenance or asset replacement can then be planned in advance if the asset condition is classified as being near to failure. Planning preventative maintenance increases network availability, reduces disruption and improves passengers’ experiences due to fewer unplanned network possessions.
Passenger Information Systems
Real-time disruption means that travelling passengers need accurate real-time information about the arrival and departure times of services, with the level of accuracy becoming increasingly important during the countdown to the final arrival or departure time. Inaccurate information will lead to ill-informed decisions about travel and reduce customer satisfaction.
If a passenger information system reports a late service, and a passenger uses the time to grab a coffee and upon returning to the platform finds the service departed on time, that passenger will be less satisfied than the passenger who decided to wait on the platform despite the forecast delay. In fact, having been told their train was late, the passenger on the platform may be pleasantly surprised when the train arrived on time; despite only receiving the service level they should expect for every journey.
The main objective of a passenger information system is to provide accurate forecast arrival and departure times and control the content and flow of information to passengers as the countdown to service departure time reduces. Stepwise modelling of train movements between train berths, complemented with weighted logic trees would appear to be an obvious way forward to inform the passenger information system of arrival and departure times. However, real-life train operation is messy, not logical, and real-time decisions involving people can be inconsistent and sub-optimal, not least because different people with different perspectives, experiences and unconscious biases make different and inconsistent decisions.
The challenge of improving passenger information is not dissimilar to that of resilient timetable planning. Historic data patterns and interventions taken either manually or automatically can be used to train the AI system to forecast train arrival and departure times. The Soft AI solution can be programmed with the pattern recognition and forecasting algorithms that, once complete, translate the real-time pattern of train services at any moment-in-time to the most likely arrival and departure times at stations.
Making a Difference to the Passenger Experience with Augmented AI
While Soft AI can radically alter how work gets done and who does it, the technology’s greatest near-term impact will be in complementing and augmenting human capabilities to contextualise and improve an individual’s journey experience. This is a real possibility – the technology is available now.
Each passenger’s experience of travelling via the railway and their level of satisfaction is individual to them. Near identical experiences in terms of punctuality, crowding and the availability of train and station facilities can lead to wholly different levels of satisfaction based on in-the-moment context, personal circumstances, individual needs and mood.
So how can context be captured and understood. Secure mobile apps are now available for passengers to enter information and feedback, voluntarily, including individual needs, mood, wellbeing and experiences. Combining these data sources with real-time 3rd party data such as connecting train, bus or taxi services and bookings, will allow communication and service interventions to be targeted at an individual level at any point before, during, or after their journey. Even when services are running to the plan and train services are good, updates would most likely be welcome to provide reassurance. Gathering the right information would allow operators to identify ‘Wouldn’t it be good if ….’ personal interventions.
With individual passenger context data available, the opportunity to truly understand an individual’s end-to-end journey experience in real-time can identify and guide tailored individual experience improvements, also in real-time. This feat is now possible by assembling available Soft AI components into viable AI solutions. Human and automated engagement can then be planned and delivered in real-time, whether services were disrupted or not. Achieving this really would be putting passengers first and making every second count.