AI in railway signalling

[Failed Unit]

Was thinking again tonight after another evening of GN being unable to restore the service. Would AI be able to help. People say that controllers often can’t think of every scenario when a crisis situation happens. That is fair I am sure a controller wouldn’t chose to skip stop a service when it going to lose more time because a slow train is let out ahead of it.

This seems like a good for AI so these decisions can be tested before implementation. Even with the automatic route setting could on-time trains be held a junction to let a late one go past (maybe recover the delay). Or let a late train get later if a driver and /or train were approaching the end of the diagram.

Even the best signaller can’t really tell what difference losing 5 minutes at Welwyn North is going to make on a London - Inverness train. AI could at least learn from previous days / masses of running data.

 

[SynthD]

How would you verify that AI was correct in anything it said? Especially as the training data has variations that need context to be understood for what they are. AI will be optimistic and miss things out or add something new to fit a pattern.

 

[devon_belle]

There may be some machine learning applications to the problem, especially if someone could write assistance software using appropriate algorithms (not necessarily Large Language Models like ChatGPT etc) to help. However, most of this technology has been around in some form since the 1990s or earlier, so I presume it would have been done by now if there was a genuine need. I don't think you'd need a massive data centre, but you would certainly want something more sophisticated than a prompt input to a commercial LLM. How do you cast the problem in a way the machine can understand? How do you provide the data? What solution do you actually want?

Sure, rail disruption is a multifaceted problem with a huge number of variables, which machines are better than humans at working out. But given that most consumer AI tools are simply word guessing machines that rely on pattern recognition and training on vast data sets, the unique and unseen nature of most railway disruption may be a case in which the tool struggles to beat humans at solving. I would bet that a more prudent solution to the inability to recover from disruption would be to redesign the timetable to reduce knock-on effects or take preventative action on sources of disruption. That's probably a £££ problem, but a better use of money than sinking it into a predatory AI start-up trying to get in before the bubble bursts.

I'd rather keep human expertise on this one. Plus, we know how it would go: the companies would get a tool that gives a small boost in efficiency/accuracy and use that to reduce their workforce rather than improve their net output.

 

[zwk500]

Algorithims are already being used in some traffic management (regulation/control) applications, and Traffic Management systems can be integrated into signalling applications that use Automatic Route Setting (ARS) if suitably compatible.

(As an aside, this would be the 'TM' bit of ERTMS refers to the traffic management system, and ETCS is the actual signalling (Train Control) sub-system).

Before everybody jumps on the AI bubble as the saviour of railway performance, there are important lessons from the Airline Industry on how AI-type systems can fall over catastrophically:

(link to YouTube describing the complete meltdown of Skysolver software for the US budge airline SouthWest).

 

[The exile]

From what I experience every day, the accuracy of timely passenger information has gone down dramatically since the introduction of ARS. Probably just a lack of consideration of passengers at the point of programming - but a good example of the fact that technology sometimes creates as many problems as it solves.

 

[stevieinselby]

From what I experience every day, the accuracy of timely passenger information has gone down dramatically since the introduction of ARS.

Is that because nobody can be ARS-ed?

 

[zwk500]

Is that because nobody can be ARS-ed?

Some versions were introduced as Automatic Route Setting Equipment. I think that got changed quite quickly...

 

[stevieinselby]

Some versions were introduced as Automatic Route Setting Equipment. I think that got changed quite quickly...

We need to get to the bottom of who came up with that one...

 

[AngusH]

You don't actually need LLM based "AI" (ie what has now become "AI") for costing available options.

Costing various options is more like the "AI" in video games, which is a much safer, predictable and reliable system.



One way is just a computer program that builds a model of each option and calculates the "cost" of each.
The system then either picks the best or displays the list to a human to choose.

(Which is pretty much like railway ticketing sites really)

You just need a suitable way of costing. Delay repay payout would be a possibility or delay minutes or some synthetic value.

If designed right it is entirely predictable, repeatable and can be debugged after the event.

Plus costing algorithms can run on your phone and require next to nothing in the way of computing power these days,
while LLM "AI" is expensive.

 

[edwin_m]

Some versions were introduced as Automatic Route Setting Equipment. I think that got changed quite quickly...

The term certainly wasn't in official use when I was on the support team for Liverpool Street, York and Yoker in the late 1980s. It may have been for the trial installation at Three Bridges, before my time.