• Our booking engine at tickets.railforums.co.uk (powered by TrainSplit) helps support the running of the forum with every ticket purchase! Find out more and ask any questions/give us feedback in this thread!

Getting everything in the right place after incidents

Status
Not open for further replies.

Wombat

Member
Joined
12 Jul 2013
Messages
299
On Tuesday I think there were a couple of incidents on the South West part of the network, which caused delays and cancellations for a time. I was lucky enough to arrive at the station in time to catch a delayed train, and it got me thinking about the process of getting trains and staff in the right places.

Many years ago I did a maths degree and one of the subjects was linear programming, which I vaguely recall to be the study of optimising systems to get things in the desired place for the lowest cost/time. While my studies never got anywhere close to the complexity of a rail network, it occurred to me that the general principles might be applicable to the problem.

It would be horrendously complicated - you'd have to consider speed limits, bottlenecks, infrastructure, driver knowledge and a whole bunch of other stuff. I wouldn't even know where to start, but it feels (in my ignorance) like the sort of problem that can lend itself to mathematical optimisation. So my question is: how does the industry go about disentangling the network when it snarls up? Is it a process of feeding data into a computer and acting upon the result, or a team of people making judgement calls based upon their experience, or a combination of both?

I'm genuinely curious. While I wasn't a great mathematician I now work in IT, so this sort of thing interests me. (Though if I'm ever given the task of devising an automated network-disentangling algorithm I will throw myself off a building).
 
Sponsor Post - registered members do not see these adverts; click here to register, or click here to log in
R

RailUK Forums

hick

Member
Joined
10 Aug 2009
Messages
96
If only! Generally it's looking at each train one by one based on which one needs looking at first. Where is it, where does it need to be next and where does its crew need to be. If it's close to the end of the day, where should it be at the end of the day, does it or the crew run any last services and would it actually be better to put it to bed now
 

TheEdge

Established Member
Joined
29 Nov 2012
Messages
4,489
Location
Norwich
Its people with experience in control looking at the situation and working out what to do and how to do it. An experienced controller in the chair during disruption is really noticeable to both staff and passengers.

Its helped by contingency planning, basically based on "line blockage" between A and B, or B and C and so on. So if that is used certain services will be cancelled. Its what causes the same services to get cancelled every time there is disruption, its the best way to deal with it.
 

infobleep

Veteran Member
Joined
27 Feb 2011
Messages
12,558
Its people with experience in control looking at the situation and working out what to do and how to do it. An experienced controller in the chair during disruption is really noticeable to both staff and passengers.

Its helped by contingency planning, basically based on "line blockage" between A and B, or B and C and so on. So if that is used certain services will be cancelled. Its what causes the same services to get cancelled every time there is disruption, its the best way to deal with it.
South West Trains said in a web chat they try to not disrupt the same passengers during different disruptions. In order to do that they can't cancel the same trains each time.
 

GW43125

Established Member
Joined
8 Dec 2014
Messages
2,045
South West Trains said in a web chat they try to not disrupt the same passengers during different disruptions. In order to do that they can't cancel the same trains each time.

What I'm told on SWT is they have a few which go before anything else when it goes up the wall near London, these tend to be:

  • Dorking services (Can take a Guildford via Epsom and change for SN)
  • Weybridge via Hounslow services (buses are kept at Staines so can set up a Staines<->Weybridge shuttle, change at Staines for the Reading/Windsor services)
  • Anything west of Salisbury turning back at basingstoke to maintain pathing on the single lines West of Salisbury
  • Then looking into cutting back Chessington and Hampton Court services and let the buses deal with them
  • Similarly, the next step involves taking out the Woking stoppers and putting additional stops on Basingstoke & Alton services
 
Last edited:

AndyPJG

Member
Joined
29 Jun 2012
Messages
420
What I'm told on SWT is they have a few which go before anything else when it goes up the wall near London, these tend to be:

  • Dorking services (Can take a Guildford via Epsom and change for SN)
  • Weybridge via Hounslow services (buses are kept at Staines so can set up a Staines<->Weybridge shuttle, change at Staines for the Reading/Windsor services)
  • Anything west of Salisbury turning back at basingstoke to maintain pathing on the single lines West of Salisbury
  • Then looking into cutting back Chessington and Hampton Court services and let the buses deal with them
  • Similarly, the next step involves taking out the Woking stoppers and putting additional stops on Basingstoke & Alton services

SWT have a quick fix for the Hampton Ct service, they can run it as a shuttle Hampton Court <> Thames Ditton <> Surbiton Plt 4 as the Dn Hampton is bi-directionally signalled. Easy fix that removes 2 tph from Waterloo.
 

aleph_0

Member
Joined
15 Sep 2010
Messages
171
If only! Generally it's looking at each train one by one based on which one needs looking at first. Where is it, where does it need to be next and where does its crew need to be. If it's close to the end of the day, where should it be at the end of the day, does it or the crew run any last services and would it actually be better to put it to bed now

That really surprises me. Whilst I'm sure an experienced employee does an excellent job, and the computational difficult of constraint problems is such that human intuition is competitive with computers, there must be room for computers to help. Timetable planning software could/should be adapted or built with this use in mind.

Based on live feeds of train position data, train crew data, the situation could be modeled, and it would be possible for hundreds of solutions to be modeled, including complex plans that a human wouldn't be able to anticipate. I'm sure the general principal would still be the same - thin out services, use the slack and capacity to get things back in order, but I would anticipate such an approach to be more successful than a purely human-driven effott.
 

GW43125

Established Member
Joined
8 Dec 2014
Messages
2,045
SWT have a quick fix for the Hampton Ct service, they can run it as a shuttle Hampton Court <> Thames Ditton <> Surbiton Plt 4 as the Dn Hampton is bi-directionally signalled. Easy fix that removes 2 tph from Waterloo.

Very good point. Similarly there's the option with Epsom line/Chessington services to tip out at Raynes Park and run empty to WPD and reverse, though a downside of this is that it would monopolise Wimbledon station throat as it comes back out.
 

Carntyne

Member
Joined
8 Jul 2015
Messages
882
That really surprises me. Whilst I'm sure an experienced employee does an excellent job, and the computational difficult of constraint problems is such that human intuition is competitive with computers, there must be room for computers to help. Timetable planning software could/should be adapted or built with this use in mind.

Based on live feeds of train position data, train crew data, the situation could be modeled, and it would be possible for hundreds of solutions to be modeled, including complex plans that a human wouldn't be able to anticipate. I'm sure the general principal would still be the same - thin out services, use the slack and capacity to get things back in order, but I would anticipate such an approach to be more successful than a purely human-driven effott.

It's a great idea, but there's no real replacement for an experienced controller during disruption, it's where they earn their money.
 

Roast Veg

Established Member
Joined
28 Oct 2016
Messages
2,200
It's no real replacement right now, but it's a very computable problem that may well have an implementation in the future. In fact, I could have done it as a dissertation to my own Computer Science degree, had I not opted for something else.
 

rf_ioliver

Member
Joined
17 Apr 2011
Messages
855
Many years ago I did a maths degree and one of the subjects was linear programming, which I vaguely recall to be the study of optimising systems to get things in the desired place for the lowest cost/time. While my studies never got anywhere close to the complexity of a rail network, it occurred to me that the general principles might be applicable to the problem.

Actually you wish it were linear...most of these problems tend to be non-linear with some really nasty higher-order effects. If you can simplify to be linear and get some sort of sensible result in a [mathematically] short amount of time then yeah!! :)

It would be horrendously complicated

Simple problems in P if you're lucky, usually NP or NP-complete and quite probably PSPACE or much much worse. Even if you're P comlex then the O(n2) is typical after addition of a few variables.

TL;DR ... we can solve it, but the Universe a) might not last long enough to run the algorithm and worse b) might not be big enough to store/process the algorithm...

t.

Ian
 

yorkie

Forum Staff
Staff Member
Administrator
Joined
6 Jun 2005
Messages
67,439
Location
Yorkshire
It's no real replacement right now, but it's a very computable problem that may well have an implementation in the future. In fact, I could have done it as a dissertation to my own Computer Science degree, had I not opted for something else.
If you can program in all the requirements for where the units should finish at the end of the day, and the route knowledge of Guards & Drivers, route availability, the demand from displaced passengers (which would vary) and various other factors, then maybe...

I'd imagine it would take a very long time to model all that accurately and the model would change as circumstances change.
 

Roast Veg

Established Member
Joined
28 Oct 2016
Messages
2,200
Absolutely, but it definitely isn't intractable - all of those variables have hard and fast rules that make up a set of heuristics, which actually make it faster to compute since there are fewer possibilities. I'd give it a go myself if I had the data, sounds like s fun challenge.
 

TheEdge

Established Member
Joined
29 Nov 2012
Messages
4,489
Location
Norwich
Absolutely, but it definitely isn't intractable - all of those variables have hard and fast rules that make up a set of heuristics, which actually make it faster to compute since there are fewer possibilities. I'd give it a go myself if I had the data, sounds like s fun challenge.

If you could do it I'm sure the TOCs would bite you hand off, but I honestly don't think in this situation much will beat an experienced traincrew manager or controller, or both.

Take traincrew managers. My depot has really good ones but a couple are brilliant, they can juggle traincrew while thinking hours/days ahead. Could a computer do that dynamically? One of ours will take crews off unaffected services, chuck them onto disrupted services, knowing he can then shuffle crews all over the place to cover the service, get someone in on a rest day at short notice for 30 pieces of silver to cover another train and then get driver Jones who is exceedingly late home in time to mean he can get his statutory 12 hours rest before his shift tomorrow. Could a computer replicate that sort of thinking?
 

Roast Veg

Established Member
Joined
28 Oct 2016
Messages
2,200
The short answer is yes. You, and others, have aptly described the level of difficulty we are talking about, but because we are dealing with relatively hard and fast truths and rules, none of this is un-modelable. This isn't a system designed to arbitrarily replace a job, rather to make that job faster and more effective - those currently employed in this do some sterling work already, but the time it takes them to do it can be reduced with the help of a well designed assistance tool. Since catching up with delays lowers the fines for late running, the business case for the TOC is already there (unsurprisingly).
 

daikilo

Established Member
Joined
2 Feb 2010
Messages
1,622
The biggest problems with incidents is that they hardly ever happen when everything is running to time and rarely are two identical. This means that before you can do any calculations you need to update the base case.

In theory, if all the data was in real time then there would be no difference, but when I xas in control on the ECML, it wasn't. We'd find that a driver was not on his normal schedule and would be turning back at a station off our patch with the result that he would agree to continue if he was delayed. Sometimes we'd be lucky and there was a driver on the cushions doing a positioning move.

Also, there are many unknowns such as when the incident will be over, knock-on effects on other patches such as trains being cancelled due to no incoming guard.

The trains themselves aren't really the issue as it is usually possible to get those requiring maintenance to where they need to be at some time, and they don't mind being 3 hours late or taking a route they'd never been on before, it's the crews, especially as now when there are relatively few spares.

What you can do is take an adapted plan and run scenarios to see which minimises risk, but you can only do that if you have someone available to run them when in reality you want all hands on the deck to keep as much running as possible.

Finally, when the incident effectively leads to a bottle-neck, you have to decide what to allow through and what will anyway so far off plan that it is better to cancel and restart further down the line by finding a train from somewhere (maybe by canceling a train in the other direction) and using the planned crew.
 

Roast Veg

Established Member
Joined
28 Oct 2016
Messages
2,200
Sometimes we'd be lucky and there was a driver on the cushions doing a positioning move.

Ah now here we would have a real sticking point for automation. This is the first thing I hadn't considered that isn't discrete in the slightest. As soon as the crew are off the books it's totally unknown and would require a human being to be in the know, as is so at present.
 

takno

Established Member
Joined
9 Jul 2016
Messages
5,040
Ah now here we would have a real sticking point for automation. This is the first thing I hadn't considered that isn't discrete in the slightest. As soon as the crew are off the books it's totally unknown and would require a human being to be in the know, as is so at present.

Lots of computerized scheduling fails because of soft factors - a decent manager will have an idea for their staff of whether they are going to be willing to take on OT, ongoing commitments which will make it difficult for them to work late and a whole bunch of other things which nobody is ever going to input into a computer just in case (they share them with other people because doing so is social, not just in case). None of these things can be reasonably heuristically modelled.

There are plenty of places on the railway where intelligent systems can offer great decision support - physical objects you can predict the behaviour of and tag monitors onto like trains are fine, and crowd behaviour can be fairly well modelled. As soon as you get to individual wet-ware though things get a lot harder.
 

mickulty

Member
Joined
8 Oct 2016
Messages
34
Taking in a colossal amount of data like what's going on with a busy rail network then making a decision as to how best to address it sounds to me like a classic use case for an "AI" algorithm with a "neural network" design, and the squishy organic versions of neural networks benefit from creative thinking and are available at very reasonable rates considering.

Afterthought: There's probably a pretty strong case for shared autonomy of some sort though? A computer may not be able to answer the question "How do we fix this?", but would probably be able to answer "What if we try this?" easily enough. But I can't pretend to know anything about the process it'd hypothetically assist.
 
Last edited:

Suraggu

Member
Joined
23 Oct 2013
Messages
964
Location
The Far North
As my role is actually being charge of running the train service it is a lot harder than people think.
The main points I can relate to restoring the train service are....

:- Traincrews being in the incorrect locations to restart a service.
:- Traincrews being out of hours.
:- Traincrews having limited route knowledge and limited traction knowledge.
:- Rolling Stock sat in depots and yards when it is needed at stations to restart services. During disruption at big terminus and stations Network Rail would rather have the platforms clear.
:- Rolling Stock available could not be compatible to run the restarted service required.
:- Rolling Stock may not have the required exam mileages to complete the amended diagrams for the rest of the day.

Because of these variables restoring the full service is extremely difficult and most of the time after serious disruption a thinned out service is run for the remainder of the day.

It's not as straight forward as following a set plan because it never happens. No two disruptions are ever the same.
 

rf_ioliver

Member
Joined
17 Apr 2011
Messages
855
The short answer is yes. You, and others, have aptly described the level of difficulty we are talking about, but because we are dealing with relatively hard and fast truths and rules, none of this is un-modelable. This isn't a system designed to arbitrarily replace a job, rather to make that job faster and more effective - those currently employed in this do some sterling work already, but the time it takes them to do it can be reduced with the help of a well designed assistance tool. Since catching up with delays lowers the fines for late running, the business case for the TOC is already there (unsurprisingly).

Hard & fast rules != heuristics

Honestly, the number of variables and the fact that some of them are most likely non-linear *and* that most of the variables are certainly non-independent means that this problem is not tractable in any meaningful sense.

If you can reduce it down to, say, 10 or so, independent, linear variables then give it a go, otherwise you're looking at a state space which has far too many dimensions to be sensible with points in that space not being "nice" integers or reals but some horrible, twisted vector (or worse).

I do this kind of stuff for living (optimisatons, system modelling) - we're normally happy if we can write down the abstract model of the system and identify potential relationships between structures, let alone actualy start writing the maths to solve this.

Just for a start, start describing the relationship between, say, trains and crew - resource allocation and timetabling are classically hard problems. Then set bounds on what a working system is and then place that system outside of those bounds and see what happens ... chaos theory might be a good place to start. Most of the time to solve these you tend to play game theory to get an idea of how to proceed, if not formally specified however (I can#t imagine Von Neumann or Turing in a railway control room, though they'd have loved the challenge!)

If you do solve this then a) there's a Nobel Prize waiting in economics, b) a Field's Medal and c) the NSA would like a word with you about solving some computationally intractable problems :)

t.

Ian
 

aleph_0

Member
Joined
15 Sep 2010
Messages
171
I agree with rf_ioliver in that these problems are computationally difficult, and I'm sure as Carntyne highlights, an experienced controller is highly valuable.

It's worth pointing out most of the data points we're talking about (where crew are, crew schedules, train maintenance cycles, overtime willingness) isn't hard to get into a computer-readable format.

The difficult bit is the problem solving bit. However, I do feel there would be room for a computational aid. An obvious quick win would be for plan verification - given the decision a person has made, flag up likely consequences/problems, or real-time changes. I know these are things that an experienced controller will be thinking of, but verification is obviously useful. There may be room for some activity to help planning - but I wouldn't want to make any strong claims here (finding the 'best' solution I expect will be too computationally expensive, checking for small variations from the plan to see if they're better would be easy, what's possible is somewhere between the two).

Ultimately, technology would/should allow the planner to focus on the bit they are good at - the problem solving, by automating out the tasks which require arithmetic rather than constraint-solving.
 

D1009

Established Member
Joined
22 Feb 2012
Messages
3,166
Location
Stoke Gifford
Isn't this what Network Rail's "Digital Railway" vision is about? It is proving to be a lot more difficult to make the systems work advantageously than the "visionaries" imagine, even before the bean counters evaluating the various competing suppliers get involved. I have personal experience of this having worked as a train planner while Network Rail was trying to update its train planning software.
 

306024

Established Member
Joined
23 Jan 2013
Messages
3,942
Location
East Anglia
Isn't this what Network Rail's "Digital Railway" vision is about? It is proving to be a lot more difficult to make the systems work advantageously than the "visionaries" imagine, even before the bean counters evaluating the various competing suppliers get involved. I have personal experience of this having worked as a train planner while Network Rail was trying to update its train planning software.

That is the theory. However NR have very little experience of the train crew side of things, and as others have said getting the crews in the right place quickly is the key to prompt service recovery.

TOCs & FOCs use various systems to document their crew diagrams, all of which would have to interface with the NR traffic management system. I'm not aware the software development has got that far yet.
 

306024

Established Member
Joined
23 Jan 2013
Messages
3,942
Location
East Anglia
......If you can reduce it down to, say, 10 or so, independent, linear variables then give it a go, otherwise you're looking at a state space which has far too many dimensions to be sensible with points in that space not being "nice" integers or reals but some horrible, twisted vector (or worse)......

Is there anything worse than a horrible twisted vector? ;)
 

bb21

Emeritus Moderator
Joined
4 Feb 2010
Messages
24,151
It's no real replacement right now, but it's a very computable problem that may well have an implementation in the future. In fact, I could have done it as a dissertation to my own Computer Science degree, had I not opted for something else.

Modelling an ideal situation and real life are two very different matters.

I speak from experience.

You are not the only one trying to work out a solution to these problems. Very little inroads have been made despite all sorts of initiatives. Most of them cannot demonstrate sufficient advantage over an experienced controller.
 
Status
Not open for further replies.

Top