Can computer modelling be used in the testing and designing of timetables on a large scale?

[The Planner]

EDIT: Perhaps someone from Milton Keynes could weigh in? Is such tech already in use and helping in train planning?

Ian knows his onions as he is ex NR, but no there isnt anything comprehensive as much as the lauded introduction of TPS back in 2010 was envisaged to solve it all. Attune will sort of automatically check compliance with TPR etc but the biggest problem is actually being able to quickly performance model. That is getting somewhere but it has a way to go.

 

[Ianno87]

AI and machine learning could fairly easily do timetabling, it will still require instructions from human operators for how much time is required at stations etc but if it has the information it needs then it could easily do it.

How does a computer pick the right outcome when there are thousands of possible right outcomes? i.e. if there's a choice between putting 1 minute of pathing time in 5 trains, or 5 minutes of pathing time in one train, which is the right outcome? Which is the most sensible? There is no one-size-fits-all answer to that.

 

[Ianno87]

It would be an obvious next step to feed this with actual train running performance, loading etc to input into modelling the next timetable (and to adjust any rules, or to identify any weaknesses, whether infrastructure, rolling stock or human, and to mitigate those).

There is the Performance/Capacity/Journey Time trade-off triangle, which are inherently in tension with each other. Trying to perfect one or two of them compromises the other, or has whole-system consequences.

 

[Ianno87]

There are many, many subjective calls. Dying people are usually very complex cases with multi-morbidity. That's why we are data led. But with subjective consideration - Even when the data says "yes we can", some people just don't want any more procedures and to die at home.

But are you thinking about what you are typing? You are seeing it as an impossibility for IT and algorithms to cacluate what it takes two people an entire day to calculate and admit that what they come up with a solution that probably isn't "correct" anyway.

You are not thinking about what you are typing. I did not state planners come up with the wrong solution, but there are multiple correct, compliant ones.

Fore example human can judge whether or not 5 minutes of pathing time in one train is a better solution than 1 minute of pathing time in 5 trains. Both are technically correct solutions, but a computer could not judge which is better in a way that is one-size-fits-all across every instance of such a decision across a whole railway system.

Right! Perhaps there's a little chip on the shoulder of some who defend "the way it's always been done" but AI is capable of some very impressive stuff nowadays. Plus train planning (as complex and a dark art as it is) has more fixed outcomes and a more easily defined "win" than some of the application of neural networks/deep learning/AI in general.

No, there are infinite fixed outcomes in train planning; the train service, stopping patterns, origin/destinations, rolling stock are all, in effect, infinitely variable to solve the problem. A human planner knows the logical tolerances that these can be varied within; a computer would either be programmed too rigidly and not see the possibilities and opportunities a commercially-minded planner would, or would come up with solutions that are too "creative" and off-the-wall.

 

[alistairlees]

There is the Performance/Capacity/Journey Time trade-off triangle, which are inherently in tension with each other. Trying to perfect one or two of them compromises the other, or has whole-system consequences.

Sounds perfect for modelling with different inputs to me.

 

[Ianno87]

Certainly the industry needs to find a way to get away from the vicious circle of:
- it takes years to do a timetable
- sometimes when a new timetable is implemented it doesn’t work, and passengers suffer (as does the revenue and reputation of the railway)
- but, as it takes years to do a new timetable, then passengers will have to suffer for years until we fix it

This completely lacks customer focus. Just because it’s the way it’s always been done definitely doesn’t mean it’s the right way.

The May 2018 Thameslink timetable was crafted over many years by human eye, with iteration and refinement. Initial introduction aside (Which was cuased by pratting about with it at the last minute), the "full" plan, crafted over multiple years development, has settled down very well, by throwing established service patterns out of the window and starting from a blank sheet of paper.

 

[Ianno87]

Sounds perfect for modelling with different inputs to me.

Easy to do I agree for a set of a few trains interacting with each other.

Doing it for the interaction between thousands of trains interwoven and interrelated with each other across the entire network playing off against each other in multi-multi-multi dimensional chess is another matter.

 

[Nean]

For those throwing words like "algorithms", "AI", "just use computers" and other buzzwords that have infiltrated industry around- I strongly suggest you ask a computer scientist about it... "AI" and "Neural Networks" need to be trained, and correctly at that- otherwise it'll just throw out any random garbage that happens to comply with the rules. Google's picture AI at one point started seeing cat faces everywhere due to the high number of cat pictures on the internet.

The fiscal reality is that any sort of system like that would cost millions to develop (we all know about government contracting- especially when it comes to IT) to still have to be manually checked by a team of humans in order to ensure the system hasn't spat out a load of garbage- if it's even able to reconcile a working timetable without human intervention.

Using 2-3 multiple computers to cross-check (as with ARS/other critical software) won't necessarily work with AI as all the systems are more than likely to turn up a different solution (almost as if you locked 2-3 different teams of planners away with no contact with one-another).

I believe that in time it will become possible, however at the moment the purported technology to fix the problem is still too nascent to be reliable and would require millions of pounds, and many years to develop (unified NHS IT anyone?)- which when people are complaining about time saved by trusting the machine is ironic.

I'm going to suffix this with: I've got no stake in this either way- I'm naturally against there being "one true way" of doing things, however get annoyed when people start throwing solutions around with no thought to implementation.

 

[alistairlees]

You are not thinking about what you are typing. I did not state planners come up with the wrong solution, but there are multiple correct, compliant ones.

Fore example human can judge whether or not 5 minutes of pathing time in one train is a better solution than 1 minute of pathing time in 5 trains. Both are technically correct solutions, but a computer could not judge which is better in a way that is one-size-fits-all across every instance of such a decision across a whole railway system.



No, there are infinite fixed outcomes in train planning; the train service, stopping patterns, origin/destinations, rolling stock are all, in effect, infinitely variable to solve the problem. A human planner knows the logical tolerances that these can be varied within; a computer would either be programmed too rigidly and not see the possibilities and opportunities a commercially-minded planner would, or would come up with solutions that are too "creative" and off-the-wall.

You just ‘train’ the computer - just like you would a human. Your argument at the moment is basically that a highly experienced planner who knows all the rules will perform better than a computer to which all the rules and options have not been supplied. Well, yes, obviously.

This isn’t about replacing train planning or train planners with some super computer, by the way. It’s about letting train planners focus much more on what the inputs should be, and analysis of the options (as well as real outcomes), whilst leaving the computing power to both do the hard work and to provide much clearer insight into what the effects of certain changes or options are.

 

[Energy]

How does a computer pick the right outcome when there are thousands of possible right outcomes? i.e. if there's a choice between putting 1 minute of pathing time in 5 trains, or 5 minutes of pathing time in one train, which is the right outcome? Which is the most sensible? There is no one-size-fits-all answer to that.

Pathing time required is something which would be specified by a human operator, using a computer to do it isn't completely replacing a human's job but just making timetabling easier and more efficient. If the computer has to decide for itself then how is a very good question which no one knows the answer to, the way AI is trained means that it is just a black box which information goes in and other stuff comes out.

For those throwing words like "algorithms", "AI", "just use computers" and other buzzwords that have infiltrated industry around- I strongly suggest you ask a computer scientist about it... "AI" and "Neural Networks" need to be trained, and correctly at that- otherwise it'll just throw out any random garbage that happens to comply with the rules. Google's picture AI at one point started seeing cat faces everywhere due to the high number of cat pictures on the internet

Completely correct, Microsoft's Tay AI chat bot had an issue in that it got trained by the public to be very offensive. How do we train an AI making a timetable? Well we probably don't as a timetable making program doesn't really use AI unless it has to make decisions itself which should be decided by a human such as the time required at stations for drivers to change, the program actually making the timetable will just be doing lots of maths to try and get the most amount of the paths wanted within the rules which the human has given the program (like time spent at stations, pathing time, etc.)

How does this relate to central control of the railways? Well if we had one program timetabling the whole railway it would be more efficient with more trains being able to run on the tracks at a time.

 

[gsnedders]

I believe that in time it will become possible, however at the moment the purported technology to fix the problem is still too nascent to be reliable and would require millions of pounds, and many years to develop (unified NHS IT anyone?)- which when people are complaining about time saved by trusting the machine is ironic.

By no means is the necessary technology too nascent, it is ultimately just an optimization problem, and optimization problems are an exceptionally well-studied field with numerous applications over the past fifty years. You don't need something following all the latest buzzwords here.

A more comparable to the railway case is that many airlines use approximate global optimization techniques like simulated annealing for flight scheduling (the choice of routes, the allocation of aircraft to those flights, the allocation of staff to those aircraft).

Ultimately I don't doubt that it is completely computationally possible today (or even two decades ago). The hard part is very much the modelling, because in the rail case especially you have huge numbers of variables.

Some of the trade-offs though make it exceptionally hard to define an objective function: things like what's the appropriate trade-off between local and intercity services, what's the appropriate trade-off between passenger and freight services, and all of these will vary across the network. (I'm much less concerned about the pathing issues, as an estimate of total passenger delay is quite possibly achievable there.)

 

[Bletchleyite]

would come up with solutions that are too "creative" and off-the-wall.

Such as, for example, Northern's disastrous crew diagramming which I believe was done by a computer? Though that was probably a case of GIGO (garbage in, garbage out) - nobody told the computer that interrelated diagrams would break things.

 

[Bletchleyite]

Doing it for the interaction between thousands of trains interwoven and interrelated with each other across the entire network playing off against each other in multi-multi-multi dimensional chess is another matter.

Computers are far better at doing that than people. It would be possible, theoretically, to model the entire network to test a timetable. Humans could introduce disruptions and see what happens.

 

[Ianno87]

Pathing time required is something which would be specified by a human operator, using a computer to do it isn't completely replacing a human's job but just making timetabling easier and more efficient. If the computer has to decide for itself then how is a very good question which no one knows the answer to, the way AI is trained means that it is just a black box which information goes in and other stuff comes out

For a human planner to be able to choose which pathing time solution is best, he/she has to have perfect knowledge of every consequence of that choice. Which leads to the planner making siloed decisions on a 'Computer says no' basis rather than seeing the whole picture if they are planning by hand, where they can build up a mental picture of the hard and soft constraints of the path they are trying to plan.

For example, a planner making a pathing time decision at Birmingham New Street may not pickup up that this affects a platform allocation at Glasgow Central, that then tightens up a single line in Scptland and creates a performance issue, if this piece of data is lost in thousands and thousands of other possible permutations.

Such as, for example, Northern's disastrous crew diagramming which I believe was done by a computer? Though that was probably a case of GIGO (garbage in, garbage out) - nobody told the computer that interrelated diagrams would break things.

The May 2018 problem waa having insufficient time for the human being element to provide the sense-check, first inputs and iterate the rules accordingly.

Remember at this point the timetable itself was fixed. Now imagine trying to define rules at very early stage where the entire parameters of the timetable require defining.

That's train crew interrelating with dwell times, interelating with journey times, interrelating with Sectional Running Times, interrelating with rolling stock diagrams, interrelating with maintenance requirements, interrelating with signaller workload, interrelating with level crossing safety, interrelating with service intervals, intrrelating with stopping patterns interrelating, with power supply limits interrelating with passenger demand, interrelating with stakeholder expectations, interrelating with access rights, interrelating with commercial aspirations, interrelating with..... You hopefully get the idea by this point.

A human being is by far the best at quickly and accurately judging the trade-offs involved and what is acceptable or not and how all the above can be varied, analysed and understood. Train planning is far more nuanced than just fitting the lines on the graph together.

The other risk is a timetable output that nobody understands. You need to understand where the risks lie in any timetable so that operational delivery teams can be briefed or action plans made (e.g. focusing on dwell times at a key station with dispatch procedures). Human eye picks these things out much better.

I'm not convinced (yet) that this is sufficiently simple or 'black and white' for machine learning to be able to identify a sensible, rational outcome. Including where a machine can identify where *not* fitting a train in (i.e. rejecting a path) is a rational thing to do even if it meets all the 'Rules' (but only just).

I am happy to eat my hat in future on this and be proven wrong!

 

[Bletchleyite]

I'm not convinced (yet) that this is sufficiently simple or 'black and white' for machine learning to be able to identify a sensible, rational outcome. Including where a machine can identify where *not* fitting a train in (i.e. rejecting a path) is a rational thing to do even if it meets all the 'Rules' (but only just).

As an aside, I think that suggests the "Rules" as you put it are actually guidelines to inform human decision. For a computer they must indeed be immutable rules (or have specified, immutable exceptions) - and so that may require some iteration and improvement of them, rather than the human pressing the override.

 

[Ianno87]

As an aside, I think that suggests the "Rules" as you put it are actually guidelines to inform human decision. For a computer they must indeed be immutable rules (or have specified, immutable exceptions) - and so that may require some iteration and improvement of them, rather than the human pressing the override.

And this is where real life gets crashed into. The Timetable Planning Rules are the formal industry document that effectively states the criteria every train path must meet to be accepted into the timetable.

For legal track access decision purposes, this becomes a fixed entitity at Version 4 each year - it would leave NR wide open to challenge if it were permanently open to machine-led evolution, e.g. "why did your software improve the rules on favour of that TOCs train, but not my TOCs?" E.g. it would stand open to exposing flaws and holes in the parameters that the system is set to work within.

 

[Bletchleyite]

For legal track access decision purposes, this becomes a fixed entitity at Version 4 each year - it would leave NR wide open to challenge if it were permanently open to machine-led evolution, e.g. "why did your software improve the rules on favour of that TOCs train, but not my TOCs?" E.g. it would stand open to exposing flaws and holes in the parameters that the system is set to work within.

And that demonstrates that the system exists in part for its own benefit (the cover your backside principle) and not for the benefit of passengers. So that pushes even more in favour of change.

In the proposal outlined as the base of this thread, the contractual issues wouldn't matter - the "TOCs" would just run what they were told.

 

[Ianno87]

And that demonstrates that the system exists in part for its own benefit (the cover your backside principle) and not for the benefit of passengers. So that pushes even more in favour of change.

In the proposal outlined as the base of this thread, the contractual issues wouldn't matter - the "TOCs" would just run what they were told.

*Sometimes* manipulating the rules is robust and the right thing to do, provided it is done transparently with all timetable participants to the scrutiny of the human eye.

For example, not having to vary a standard hour passenger service to fit a freight path in one hour.

 

[Bald Rick]

I’ve watched this thread with ‘interest’. It seems to be a simple case of people who don’t understand what train and resource planning is about (it’s not just timetabling, and it’s not just based on rules and algorithms), arguing that computers and algorithms can sort it all.

It’s clear to me that we need a simple guide to what train planning is, and a more complex one for those that are more interested. The first is in production.

There are many, many subjective calls. Dying people are usually very complex cases with multi-morbidity. That's why we are data led. But with subjective consideration - Even when the data says "yes we can", some people just don't want any more procedures and to die at home.

But are you thinking about what you are typing? You are seeing it as an impossibility for IT and algorithms to cacluate what it takes two people an entire day to calculate and admit that what they come up with a solution that probably isn't "correct" anyway.

A case in point. It will take one planner maybe 5 minutes to come up with various solutions, using some high end computing. It *can* then take the best part of the day to discuss those solutions with the people who are contracted to run the train service, as it may have implications on their passengers, resource planning, and the contract with their service specifier.

But, as an aside, I would never dream of telling anyone this their work for processing 4D models of blood flows (or whatever) is being done inefficiently or ineffectively. So why does anyone think it’s right to tell train planners the same?

EDIT: Perhaps someone from Milton Keynes could weigh in? Is such tech already in use and helping in train planning?

There have been several, including one who in my assessment is the best train planner I’ve ever met (and I’ve met a lot!)

Train planning, although has a lot of science behind it, is something of an art. I simply can’t describe how much computing power was thrown at the Thameslink 2018 timetable, for several and years, and it still needed people like @Ianno87 to sort it out.

To put some scale on the numbers, there’s roughly 24,000 trains a day on the Network, each with, on average, about 50 timing points. Then there’s the resources required to run it (different types of rolling stock with different performance characteristics, and route clearance) then traincrew who are limited to certain routes, certain types of rolling stock, and will have different rostering terms and conditions depending which company and depot they work at. Then there is the possibility of a delay incident, of about a hundred causes, between any two of those timing points, that might last anything between 1 minute and several hours. This might delay any one of the trains in that area, between any two of its timing points, for anything between 30 seconds and (let’s say) 30 minutes, which may or may not affect subsequent trains.

 

[The Planner]

Computers are far better at doing that than people. It would be possible, theoretically, to model the entire network to test a timetable. Humans could introduce disruptions and see what happens.

You can do that now and have been able to for a good number of years, and it is done on smaller scales but the computational power to do it and the time it takes is still quite onerous as i pointed out earlier. But even then that only spits out numbers on punctuality based on a number of "runs" with historic TRUST delay data put in, it cannot make a decision on its own (yet) on stepping up units, what gets regulated etc. during perturbation, it just tells you what happens and can point you in the direction of what to look at. Then the cycle starts again.

 

[lordbusiness]

Not many years ago I was fortunate enough to visit a United States Navy Aircraft Carrier. Part of the tour was the space where the movement and management of the flight deck was planned and controlled from. For those that don't know, an aircraft carrier flight deck is one of the most complicated, congested and dangerous environments ever seen.
It consisted of a large scale model of the flight deck and hangar deck. Aircraft were represented by scale models with numerous coloured metal discs on top to indicate whether the aircraft was serviceable, armed, fuelled etc. The whole operaton being controlled by an Officer and a few sailors who moved the models around in radio contact with the folks up top and the flying control team.
When we asked our guide why it was such a primitive system on a multi billion dollar state of the art warship, the reply was "The navy spent millions trying to develop a computer programme that could do it as well and as quickly as the human brain- it didn't work".
There are some things a computer can do, some it can't.

 

[Domh245]

Much as I accept the industry needs to change I fear Mr Shapps has little idea of the complexities of the railway which coupled with his way of doing business may bring further conflict with the TU's.

But then again, maybe thats what he wants?

You wouldn't expect the transport secretary to be up to speed on all the details and complexities - there's a lot that he needs to be on top of as part of his brief, and the minutiae of timetabling rules isn't one of them. He identifies the (perceived) issues, the industry then needs to work to fix them.

 

[gallafent]

[…]
It would be an obvious next step to feed this with actual train running performance, loading etc to input into modelling the next timetable (and to adjust any rules, or to identify any weaknesses, whether infrastructure, rolling stock or human, and to mitigate those).
[…]
This completely lacks customer focus. Just because it’s the way it’s always been done definitely doesn’t mean it’s the right way.

… yup … and the obvious subsequent step after your obvious next step is to add the next aspect of the data set, which is where people actually want to get from and to and when, if they weren't constrained by the existing timetable, on a regular basis, which would address your second point I quoted here … and then to model the non-regular demand …

Both those are a lot less trivial than the basic constrained timetabling problem we started from, and used to be impossible or at least intractable, but we are probably moving quite fast towards being able to model and solve this sort of constrained problem with ML systems. (I'm not an expert in the field but sometimes read (the abstracts, usually, of) academic papers that are already starting to do this sort of thing, e.g. complex physical systems simulation …)

 

[Roast Veg]

Disclaimer: I am a qualified Computer Scientist.

We are still improving and refining models for machine learning, and it's becoming obvious over time that any task can have its human element at least partially improved. Instances where this has failed (aircraft carriers, crew rostering, and other such examples) are cases of poor specification. To say it is impossible for a machine to devise the timetable is incorrect, categorically. It might bankrupt the country in the process of attempting it, though.

 

[baz962]

I'm loving this particular thread , but by crikey it's making my head hurt. Where's my railway rule book , I need some light reading. As an aside , if you are ever in San Diego , California , there is a us navy aircraft carrier that you can tour around.
Sorry for going off topic.

 

[Ianno87]

Disclaimer: I am a qualified Computer Scientist.

We are still improving and refining models for machine learning, and it's becoming obvious over time that any task can have its human element at least partially improved. Instances where this has failed (aircraft carriers, crew rostering, and other such examples) are cases of poor specification. To say it is impossible for a machine to devise the timetable is incorrect, categorically. It might bankrupt the country in the process of attempting it, though

And to specify correctly, you need to carefully elicit thousands and thousands and thousands of direct and implicit requirements from a set of the finest timetable planners in the land from Network Rail and the TOCs (who, at the end if the day are timetable planners, generally not computer scientists). And even then you won't get everything.

And these timetable planners are usually to be found engaged full time with planning the timetable!

 

[route:oxford]

And to specify correctly, you need to carefully elicit thousands and thousands and thousands of direct and implicit requirements from a set of the finest timetable planners in the land from Network Rail and the TOCs (who, at the end if the day are timetable planners, generally not computer scientists). And even then you won't get everything.

And these timetable planners are usually to be found engaged full time with planning the timetable!

When it gets to the stage of the argument where claims are made that progression is impossible because the people who are currently undertaking the task are completely overwhelmed by their workload, then it's time to admit the current method is no longer fit for purpose.

 

[Bletchleyite]

You can do that now and have been able to for a good number of years, and it is done on smaller scales but the computational power to do it and the time it takes is still quite onerous as i pointed out earlier. But even then that only spits out numbers on punctuality based on a number of "runs" with historic TRUST delay data put in, it cannot make a decision on its own (yet) on stepping up units, what gets regulated etc. during perturbation, it just tells you what happens and can point you in the direction of what to look at. Then the cycle starts again.

That's something that could potentially provide big benefit in the future. With more and more people purchasing travel via journey planners (and by and large sticking to it even if they don't have to) that gives you a huge amount of data to play with in working out who's likely to be making what connections and the likes, and being able to, for example, model the possible effect of holding a connection in deciding if it's best to do that, or even to know in advance you're going to need some taxis (so get the guard on affected trains to go and find out exactly how many are needed, and provide that to the system, which could then automatically order them from a partner taxi company).

At the moment we mostly don't hold connections as it's too complex to work out the knock-on effect, so the assumption has to be that unless it's a really simple case like one of the Cornish branches where you know you could hold for the length of the turnaround at the terminus minus the end-changing time you don't hold. Modelling could provide the knowledge to make a better decision - which may still be no, but it'll be yes more often than it is now.

New thread to discuss this side of things further: https://www.railforums.co.uk/thread...-decisions-e-g-whether-to-hold-trains.205883/

 

[Ianno87]

When it gets to the stage of the argument where claims are made that progression is impossible because the people who are currently undertaking the task are completely overwhelmed by their workload, then it's time to admit the current method is no longer fit for purpose.

Didn't say it was impossible. But there'd be a severe drain on timetable planning capability whilst it was done (which would have the consequence of inhibiting the level of tolerable timetable change) for a good number of years. And *even then* fron what we know and understand about timetable planning there it is far from certain of achieving the desired outcome (for the reasons I and others have taken to the time to spell out upthread) even with bottomless pits of money thrown at developing such a system.

Again, I repeat, I do not think it is necessarily impossible. But it shapes up to be a piece of software programming that makes beating a Chess or Go Grandmaster look like a GCSE project. How much money would *you* be prepared to stump up for this, with no certainty of success?

 

[Bletchleyite]

Again, I repeat, I do not think it is necessarily impossible. But it shapes up to be a piece of software programming that makes beating a Chess or Go Grandmaster look like a GCSE project. How much money would *you* be prepared to stump up for this, with no certainty of success?

It looks like the kind of thing which would be a good collaboration with a Computer Science department at a University, which would substantially reduce the cost.