Artificial – rather than human – intelligence: an existential threat to transport modellers?

A DALL-E generated AI image of “a painting in the style of van Gogh of five people having a conversation”

A DALL-E generated AI image of “a painting in the style of van Gogh of five people having a conversation”

Is 2023 the year that transport planning embraces AI? There can be hardly any transport planner who hasn’t heard of, or had a go at using ChatGPT, the free OpenAI chatbot.

David Metz described his experiences recently in an article in Local Transport Today (LTT 872), and concluded that there isn’t that much to learn yet from AI (and perhaps that’s not surprising as all ChatGPT knows is what it has been fed in its training dataset, and even that is time-limited to a cut-off date of 2021).

Landor LINKS will organise a specific artificial intelligence and transport event, Transport AI, on 5 October and, of course, AI will feature in our Modelling World 2023 programme, when we meet in Birmingham on 2 and 3 November

ChatGPT is a large language model with apparently 175 billion parameters, which is inevitably a lot more than a behavioural model as we use in transport modelling, with a few parameters estimated statistically from a much smaller dataset. That’s machine learning for you.

And of course many of the examples you have seen, and benefits for us as transport modellers may be the AI writing yet another report or proposal (and yes, I have heard anecdotally of winning bids having been written by ChatGPT).

But its capabilities are wider: when I needed the parameters of a lognormal distribution to be estimated from a bunch of numbers, one colleague wrote a clever Python program, another asked ChatGPT - the values they came up with were within a few percentage points of each other. 

Someone else translated a small program from one computer language to another – having said that, everyone I spoke to warns that anything created by ChatGPT needs careful checking, as you would with any new entrant to our profession. David Metz gives an excellent example in his LTT text.

I caught up with four people that I consider fast followers in the artificial intelligence space:

• Philippe Perret, Strategic Consultancy Director at Amey with a history in mobile phone data analytics at CitiLogik

• Anna Jordan, who co-founded Alchera Technologies six years ago and is currently their CEO

• Andrew Browning, whose AI driven product Schemeflow aims to generate draft Transport Statements, Transport Assessments and Travel Plans for transport consultants in minutes rather than days

• Laurence Chittock, a Transport Modelling Project Manager at PTV, specialising in EV infrastructure planning

One thing that becomes immediately obvious in our discussions was that, although I use the term Artificial Intelligence liberally, machine learning is probably more appropriate for most of the current applications in transport modelling.

Says Andrew: “AI is a very loose term. It’s almost like how medieval people thought about magic. Anything that they couldn’t understand and explain was seen to be supernatural (either magic or God). Today we have the same with AI - if I can’t understand how a computer could do this with rules-based logic, it seems like AI to me!”.

Laurence concurs: “Some approaches that are currently called AI are really not much more than smart algorithms or data processing techniques. I would also separate what I call qualitative AI, large language models like ChatGPT, from quantitative AI, providing predictive traffic management insights that generally outperform traditional rule-based systems”. 

Trust is an issue that comes up regularly. Of course, for those that distrust the modelling we already do, with mathematical relationships estimated from observations, and using statistical methods - looking for causation, rather than just correlation, AI will prove to be even more of a black box. 

Having said that, it should not be impossible to carry out sensitivity tests of any AI predictions, and understand better which inputs affect outcomes most, just as we should with those models derived more traditionally.

For Anna it's the loss of accountability that worries her most. This has always been a concern with self-driving cars, CAVs, and I hope you have at some point come across the trolley problem (and if not, see this quick introduction. If reports are written (at least partly) by AI and its sources are dubious or non existing, or its algorithms biased, who is responsible?

If the machine learning's analyses in your report are wrong, where lies the blame?

Anna: “Regulation and clear guidelines will play their part in how the dust settles, but for now the advice would be: use with caution”. Laurence agrees: “’Even if in a machine-learning context the modelled outcome is the same as for a more traditional calculation, ultimately the answer to the question: “Why?” would be ‘because that’s what our system learned is the most probable outcome’”. Says Philippe: “this is really a case where the computer says ‘No!’. The critical mind of a human analyst cannot be replaced”.

Bias cannot be ignored: the rubbish in, rubbish out mantra will not disappear. ChatGPT has been accused of being both left- and right-leaning, and even prejudiced. It doesn't provide references and David Metz discovered that some of its statements and statistics are not supported by real sources (but that is a problem not limited to artificial intelligence, of course).

Again, that does not have to be a deal-breaker. Philippe makes the valid point in this respect, that rather than relying on its training dataset of unknown provenance, it is possible to feed a chatbot with controlled, internally created data, that should avoid such bias.

And Anna adds: “Having the right data input is key, and far more affordable and easy to implement than people expect. Using AI without having a solid understanding of your data real-estate, what is trustworthy and reliable data (and holding your suppliers to account!) is fraught with difficulty, and likely to fail.

Purpose-built digital infrastructure products have matured well beyond what might originally have been an unmanaged data lake on the cloud. The public sector needs to take an active role in scoping what that digital infrastructure should look like for their regions. It makes it far more cost effective to then be able to use AI (again and again) - it's a virtuous circle.”

If you feel that AI is a solution looking for a problem, the experts are less concerned. Some of the applications might be found in unexpected ways. Laurence suggest that it should be possible to allow users to query our software and models with natural language questions (rather than a sequence of buttons or code).

For example ‘Please show me the areas in the model with the worst bus accessibility and recommend how this could be improved’. This would have the potential to open up modelling to a wider audience, helping with transparency, diversity and also trust.

This immediately raises for me another concern - will such reports and analyses become more and more uniform or bland, or will we as human experts take the opportunity to use the released time to look for and report the nuances, outliers or new-to-be-gained insights? 

Anna is confident: “The scale of the raw data that is now available, and needs to be considered lends itself perfectly to machine learning, pulling out the relevant patterns - letting a transport professional make a more informed decision, taking away the grunt work”.

And Philippe adds: “For me, AI and machine learning more specifically, really shine in pattern recognition, for example when creating travel pattern insights from mobile network data”. I have heard the argument that, rather than using the term machine learning, we should talk about machine teaching.

Andrew concludes: “Why we are having this conversation now is that the last 12 months have marked a big shift from closed (needing a team of PhDs who understand statistics to do machine learning – and which may cost you £10Ms to do if you are doing it seriously) to open (now you can play with or just plug into other people’s models and pay a small fee to do so).

Tools like Chat GPT and Midjourney have popularised the use of AI because they are a) free and b) easy to use”. And Philippe adds: “The barrier to entry is really low. For anyone with some Python knowledge, scikit-learn is a free machine learning library designed to interoperate with the Python numerical and scientific libraries NumPy and SciPy. It can be used by anyone and there are many examples and tutorials online. It is one the main packages we use ourselves for analysis”.

I could not resist, and asked ChatGPT: “Is Tom van Vuren a famous transport modeller?”. The answer:”As of my knowledge cut-off in September 2021, I couldn’t find any widely recognised information about a person named Tom van Vuren as a famous transport modeller”. Ouch!!

Tom van Vuren is the Policy Director at the Transport Planning Society and a Visiting Professor at the Institute for Transport Studies at the University of Leeds. He recently joined Amey as Strategic Business Partner