Journal of Scheduling 6: 437±455, 2003.

# 2003 Kluwer Academic Publishers. Printed in the Netherlands.

A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS

ANTHONY WREN, SARAH FORES, ANN KWAN, RAYMOND KWAN,

MARGARET PARKER, AND LES PROLL
School of Computing, University of Leeds, UK

ABSTRACT
A substantial part of the operating costs of public transport is attributable to drivers, whose efficient use
therefore is important. The compilation of optimal work packages is difficult, being NP-hard. In practice,
algorithmic advances and enhanced computing power have led to significant progress in achieving better
schedules. However, differences in labor practices among modes of transport and operating companies
make production of a truly general system with acceptable performance a difficult proposition. TRACS II
has overcome these difficulties, being used with success by a substantial number of bus and train operators.
Many theoretical aspects of the system have been published previously. This paper shows for the first time
how theory and practice have been brought together, explaining the many features which have been added
to the algorithmic kernel to provide a user-friendly and adaptable system designed to provide maximum
flexibility in practice. We discuss the extent to which users have been involved in system development,
leading to many practical successes, and we summarize some recent achievements.
KEY WORDS: driver scheduling; public transport; mathematical programming; heuristics

1. INTRODUCTION
The problem of scheduling public transport and its drivers has received much attention,
particularly among the OR community, over many years. Wren (1998) has been involved since
1967 in exploring many different solution techniques which can be applied to NP-hard public
transport scheduling problems, and, in particular, to the driver scheduling problem. The most
successful commercial driver scheduling packages, for example, TRACS II (Fores, Proll, and
Wren, 2002) and HASTUS (Hamer and Seguin, 1992), model the problem using mathematical
programming.
Mathematical programming approaches are basically of two types. Generate-and-Select
techniques, including TRACS II, rely on a pre-generated set of shifts. As, for computational
reasons, this set can only be a small subset of those potentially available, such techniques cannot
guarantee to find an optimal solution to the model. In principle, column generation/branch and
price techniques (Wolsey, 1998; Barnhart et al., 1998) can overcome this difficulty by generating
``useful'' shifts from the implicitly defined full set of potential shifts, as and when required. In
practice, even using these latter techniques, the heavy computational load means that proven
optimal solutions to the model cannot, in general, be found. Fortunately intelligent use of
heuristics and increasing computer power allow, in many cases, acceptable schedules to be
achieved in reasonable times, using either generate-and-select or column generation.

* Correspondence to: Les Proll, School of Computing, University of Leeds, Leeds, LS2 9JT, UK. E-mail:
[email protected]
438 A. WREN ET AL.

More recent solution attempts have included methods which employ local search strategies to
improve current schedules. Kwan et al. (2001) describe the use of genetic algorithms and suggest
how they can overcome the limitations of a totally mathematical programming approach in
solving ever larger problems. Smith et al. discuss how constraint programming can be used to
reduce the problem size and aid subsequent attempts to find a schedule. Shen and Kwan (2001)
describe a tabu search heuristic for constructing shifts. Despite some successes, none of these
approaches has been shown to improve consistently on mathematical programming over a wide
range of problem scenarios.
Most of the theoretical aspects of the TRACS II system have been presented elsewhere. The
purposes of this paper are to show how theory and practice have been brought together in a
user-friendly and adaptable system providing maximum flexibility. We present a picture of the
whole, and discuss some of the features which have been added to the algorithmic kernel to
make it more usable in practice and to highlight some examples of its use.
In Section 2 we explain our motivation, while Section 3 outlines the driver scheduling
problem. Section 4 provides an overview of the TRACS II system, while Section 5 introduces
the user interfaces. Section 6 surveys some of the highlights of the system development over a
period of twenty years. Section 7 summarizes some uses of TRACS II, showing several of the
difficulties which have been overcome. The present state, and some possible future
developments, are summarized in Section 8.

2. MOTIVATION
The specific scheduling needs of transport undertakings vary greatly, depending on the nature of
the area being served, the type of operation (road or rail), the views of individual managements
and drivers' representatives, and historic developments. Many driver scheduling systems are
installed for users after lengthy tests and modifications lasting several months, and are thus very
costly. Our design philosophy therefore, has been to provide a modular system which requires a
minimum of change from one application to another, and which can be easily used by
schedulers with a PC, whether or not they are computer literate. To this end we have designed
TRACS II with a wide range of parameters which reflect the needs of most organizations, and
can be easily understood by newcomers. The algorithms are independent of the scheduling rules.
A large set of potential driver shifts (up to a few million) is generated early in the process, each
candidate being tested against predefined parameters. This set is refined, first by heuristics, and
then by a specialised integer linear programming (ILP) process, in order to obtain a near-
optimal working schedule. The system is driven by a graphical user interface.
TRACS II serves both as a regular scheduling tool and as a device for determining the costs of
proposed revisions of drivers' conditions or new service levels. Its benefits lie both in its ability
to achieve substantial operating efficiencies and the power it gives to management to examine
speedily alternative operating scenarios.

3. DRIVER SCHEDULING
Once a vehicle schedule has been devised, the problem of driver scheduling involves partitioning
the vehicle work into driver shifts which must be valid according to labor rules, while the total
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 439

set of shifts (the schedule) should reflect the operator's definition of efficiency. The measure of
efficiency may be the total number of shifts used, the total cost in paid hours, or some
combination of the two; it may occasionally include a measure of subjective quality.
A typical shift will constitute a day's work for a driver, and will generally include assignments
to more than one vehicle, with provision for at least one meal break. Although we refer here for
simplicity to a day's work, we include night shifts which may start late on one day and finish
during the morning of the next day. We also refer in this text to the scheduling of a single day at
a time, although this may include the option of associating some morning work with night shifts
from the previous day.
Drivers can only change vehicles at specified points. These are usually places conveniently
located to an interchange or rest area or where there is suitable transportation to or from
another such point. The location/time pairs of such places can be identified from the vehicle
schedule and are known as relief opportunities (ROs). The indivisible time periods between
successive ROs on a particular vehicle have to be allocated to a single driver and are known as
pieces of work. A shift will often include consecutive pieces of work on one vehicle followed by
work on other vehicles. The gap between such spells of work on different vehicles may be long
enough to allow a meal break, or it may simply conform to rules regarding the minimum time
necessary to change vehicles. Figure 1 shows a section of vehicle work and an example of a valid
shift that might be formed.
Many factors are involved in specifying working conditions of drivers, including a maximum
total driving time, a maximum spell spent driving without a break, time periods in which
breaks may take place, minimum duration of meal breaks, etc. These may be different for
each operating organization. Not only does the specific range of acceptable hours vary, but
particular working practices may be encouraged or prohibited, for example, by limiting the
number of shifts with particular characteristics. Labor agreements are also often different in
different countries or for different transportation types. Geographical considerations also affect
the type of shift which may be appropriate; for example, long-distance train drivers may
work on only two vehicles, out and back, while intensive urban bus operation may allow
several changes of vehicle in order to package most efficiently the work within a minimum set of
shifts.

Shift
Piece of Work
Mealbreak
}

Vehicle 1
0650 0810 0910 0950 1130 1240 1350 1500 1540 1610
D A A A B C A B B A

Vehicle 2
0750 0840 1000 1120 1230 1400 1500 1550
D C B C B B B C

Relief Time Relief Location

Figure 1. A section of vehicle work.

440 A. WREN ET AL.

Large transport undertakings often involve the use of many driver depots. While urban bus
companies frequently restrict their drivers to vehicles from their home depots, allowing the total
driver scheduling problem to be subdivided, there is often more flexibility within regional
companies, although there may be restrictions on the routes or vehicle types which may be
operated by drivers from particular depots. Scheduling of large train operations tends to be
more complex. We have successfully solved problems in which drivers have been spread among
more than 20 depots, while drivers from any one depot are only qualified to drive particular
routes and particular traction types; as any particular route can generally be driven by drivers
from more than one depot, the total problem cannot efficiently be subdivided. There may be
restrictions on the numbers of drivers available at particular depots. It may also be necessary to
constrain the number of shifts of particular types at certain depots, for example to ensure that
there is a fair balance of early and late shifts.
Drivers may have to travel from a signing-on point to the point where they are to commence
driving, or between points where successive spells of work finish and start, or from the point
where they finish driving to a signing-off point. Such travel may be on foot or by taxi (in which
case a standard time allowance is given), or by a frequent bus or train service (in which case a
standard time may also be appropriate), or as a passenger on one or more scheduled services (in
which case it is necessary to refer to the timetables of any services which may be used).

3.1. Capturing user requirements

Many, but not all, of the organizations with which we have worked have a set of written rules
setting out the agreements between management and unions, which govern the composition of
driver shifts. Some of the organizations also abide by unwritten rules. When we started
investigating driver scheduling 35 years ago, we spent considerable time with several different
bus operating authorities in order to compile a body of rules which, by suitable adjustment of
parameters, would together meet all the conditions we had met, both written and unwritten.
This was initially a difficult and complex process.
Our normal pattern in the early days of our research was to try to extract all relevant rules at
an initial meeting, and to obtain sample copies of some existing schedules. We would then
examine the schedules, looking for examples where the rules as we understood them had
apparently been violated. These examples were then raised with the company management, and
a better understanding was achieved. We would then try ourselves to create a schedule by hand,
and would present this to management. Sometimes this would lead to our being told that we had
violated some further unwritten rules, or had misunderstood some situation. Sometimes too, we
would be told that we had failed to take advantage of some hidden ways of circumventing the
given rules.
Two typical difficulties of our early days may illustrate some of our difficulties in capturing
the rules. Much of our early work was done with Leeds City Transport. About 2 years after
starting this we were able to present to management the first computer-produced schedules that
appeared satisfactory to us. We were told that we had broken some rules which had not
previously been given to us. On asking why these rules were only being revealed now, we were
told that they had not wanted to discourage us by telling us all the difficulties at once! Two years
work had to be restarted.
Another example came from Dublin in the 1970s. An important constraint in driver
scheduling governs the maximum continuous time a driver may spend in charge of a bus.
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 441

This was not written in the list of rules given to us, so we asked for clarification. We were told
that the maximum was 4 hr 30 min, but the chief scheduler added the rider that they only
exceeded this ``if necessary.'' We asked when it would be necessary, and were told that it would
be necessary whenever it enabled them to make an improvement. Under pressure they conceded
that the absolute maximum was 4 hr 45 min, but after some thought one of the schedulers
added that they hardly ever exceeded this! When we examined their existing schedules we found
an example of 5 hr 7 min continuous working time, so we set this as our maximum, adding
penalty costs to any shift with a stretch exceeding 4 hr 30 min.
Despite such difficulties, by the 1980s we had been able to put together a comprehensive list of
parameters which would often be sufficient to satisfy all the needs of a new user, although new
situations would arise from time to time. The first live user of one of our systems was London
Transport Buses, and we worked closely with LT staff over a period of 18 months to ensure that
all their conditions could be met. As a result, the system was installed successfully at the end of
1984, and is still in use by some of LT's successor companies today.
As our portfolio of successful installations has grown, so the attitudes of schedulers have
changed from a position of disbelief to one in which most potential users are keen to ensure that
the system will meet all their requirements. Nowadays good schedules can be obtained for most
new users simply by appropriate settings of a very comprehensive list of parameters. However, it
is important that we do not pressurize schedulers into accepting anything that is not completely
satisfactory. From time to time we extend the system's parameter list. TRACS II separates the
building of potential shifts from the optimization phase, and the generation of potential shifts
itself follows context-free processes. These processes invite the system to validate potential
shifts, and the validation routines can be expanded as necessary to include new features.
When we turned to train driver scheduling in 1994 we had to increase the parameter list quite
extensively, and indeed we rewrote the shift generation process completely. However, the
rail scheduling rules were normally well defined, and it was not difficult to add the relevant
new features while ensuring that all previously captured features of the bus industry were
preserved.

4. TRACS II
TRACS II is a driver scheduling system consisting of a suite of modules. It is designed to be as
portable as possible, so its main optimizing component is independent of the rules governing the
construction of any shift. Our earlier IMPACS system was installed for London Transport
Buses in 1984 and has been used as the main bus driver scheduling tool in London ever since.
IMPACS and TRACS II together have been used to construct real operational schedules in
around 100 bus and train companies, either as the regular driver scheduling tool, or to test
alternative strategies or to make once-off schedule improvements. Most recently, TRACS II has
been installed in each of the twenty-five operating companies of the UK's largest bus group,
with around 10,000 vehicles. Each of these companies has a different historical background;
some are former city-owned operations, others are regional private companies, while two are
divisions of the former London Transport. The scheduling conditions therefore vary greatly
between companies.
Essentially, TRACS II has two main phases: generation of a large set of potential shifts,
possibly up to a few million or more, and selection of a subset of these shifts which together
442 A. WREN ET AL.

form a schedule. In practice, there are several stages in the solution process of TRACS II,
controlled through a workbench or user interface (Section 5):
 data preparation, input and validation
 prespecification
 identification of travel possibilities between ROs
 generation of one or more sets of valid potential shifts
 reduction of the shift set (if required)
 merging sets of shifts
 computation of penalties
 selection of a subset of shifts which covers the vehicle work and minimizes an objective
function
 postprocessing
 output
 shift validation.

4.1. Data preparation, input, and validation

The system depends on two main data files. The first, the TRA (Travel) file, contains details
of the vehicle work to be covered (represented by sequences of ROs), together with information
about any other services which may be used by drivers travelling as passengers between relief
points. This file also contains data regarding relief points to be used, a table of standard walking
or riding times between them, allowances for drivers to sign on or sign off at each point, details
of depots and any restrictions affecting them, and other data relating to the physical system to
be scheduled. The first process of TRACS II, VALIDATE, checks the format and logic of the
data in this file.
The second main file, LAB (Labor), specifies the rules governing the formation of
individual shifts. Some of these rules are hard, in the sense that they must be followed, while
others are soft, in the sense that they are used by the scheduler to guide the system to generate a
reasonable set of potential shifts out of the many billions that might be allowed by only the hard
rules. Parameters within this file define the different types of shift which are allowable, and for
each type of shift such features as the maximum amount of continuous driving, the positions
within a shift of meal breaks, and their minimum durations, the minimum amount of work
within a shift, and the maximum elapsed time from start to finish of a shift. Some of the features
may be considered as hard parameters by some users and soft parameters by others.

4.2. Prespecification
A facility is provided whereby the user may specify in advance that certain shifts, or features
of shifts, must be included in the schedule. This may be used when special features are required,
even if they might detract from optimality. When a user specifies a complete shift, that shift need
not be legal according to the parameters specified; sometimes a scheduler knows that there is
vehicle work which cannot be covered without a dispensation to break some rules.
Partial shifts may be specified. For example, it is possible to specify the start of a shift, and/
or its finish, allowing the system to determine how best it should be filled out; this is particularly
useful when it is intended that morning and afternoon school runs should be undertaken by the
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 443

same driver. The user may also specify certain ROs when drivers must be changed, or may link
two ROs together to show that a driver finishing a spell at the first of these should then take up
another vehicle at the second specified RO. Partial specification must be such that it is possible
to construct a legal shift including the specified feature.
Although prespecification may be useful in certain circumstances, it is generally used
sparingly. Many users are happy to allow the system to produce the full schedule without their
intervention.

4.3. Identification of travel opportunities

It has already been stated that drivers may have to travel between portions of a shift, or to or
from its start or finish. In urban bus operation it is often sufficient to specify standard walking
or riding times between each pair of points, but in train operation, and in some bus operation,
drivers have to travel as passengers on scheduled services. The TRAVEL component
determines, for each RO (time and place) specified in the TRA file where the driver may
leave a vehicle, the earliest time that a driver can reach each of the other points on the system.
The journey may involve traveling on up to two scheduled services, with intermediate walking if
appropriate. Similarly, for each RO, it determines the latest time a driver may leave each of the
other points in order to take over the driving of a vehicle at the RO in question. It should be
noted that the system calculates travel opportunities for each of the ROs separately; the best
route to take between two relief points depends on the availability of scheduled services at the
time in question.
TRAVEL uses the vehicle data, together with the table of standard walking or riding times,
to compute possible passenger journeys for drivers. It is possible for a driver also to use services
which are not to be scheduled (possibly services provided by a different company), or to join or
leave a vehicle at a place which is not a relief point on the service in question. Such services and
additional points are identified in the data with codes indicating that they are not to be used
other than for passenger travel by drivers. It is also possible to specify times at which taxis might
be used in certain circumstances.
TRAVEL produces a file detailing all the many passenger travel opportunities. It is quite
normal to have over 1000 ROs, and up to 100 actual relief points, so a further file is produced
here which is used in the output (Section 4.10) to provide a detailed description of the journeys
used in the course of any shift.

4.4. Generation of one or more sets of valid potential shifts

The next stage, BUILD, makes use of the above travel information to generate potential shifts
which will later be subjected to the scheduling process itself. It considers all the possible
combinations of pieces of work which satisfy the given parameters and can be accommodated
within a shift with no more than four spells of continuous work. Although manually produced
schedules frequently have shifts with more than four spells of work, this has usually arisen
because human schedulers frequently have difficulty in fitting all the work efficiently into good
shifts; they are left with some pieces of work which they exchange with small pieces of shifts they
have previously constructed until they are able to fit everything together, often in a less than
efficient manner. It has been our experience that TRACS II has always been able to create
schedules more efficient (see Section 3) than the manual ones without exceeding four spells of
444 A. WREN ET AL.

work, and frequently without using more than three. Between any two spells of vehicle work
there is necessarily unproductive time, so schedules with fewer separate spells of work per shift
are inherently more efficient. Despite this, we do know of cases where it might be desirable to
have a few shifts with five or more spells of work; the case for this can often be identified from
properties of the data, and we envisage allowing the system in future to create a few potential
shifts with more than four distinct spells.
Since the number of possible legal combinations of pieces of work is generally enormous,
BUILD applies several filtering heuristics during the generation process to eliminate some
partial shifts that are considered to be redundant. Such partial shifts include those which are
entirely contained within other partial shifts with the same start and end ROs but covering more
driving work. There is no upper limit to the number of shifts generated by BUILD. The size of
the final set depends on the number of ROs, the frequency of ROs on any particular vehicle, and
the slackness of the parameters. A typical scheduling run might generate 100,000 potential
shifts, but cases with well over a million have been processed.
If all pieces of vehicle work cannot be covered by a BUILD run, the user has two options. The
parameters can be relaxed and the full run repeated, in which case an unnecessarily large
number of shifts may be generated. Alternatively, a new run may be undertaken in which
relaxed parameters are used to generate only shifts which contain work which was not covered
by the first run. It is sometimes appropriate to run BUILD several times with different
parameters in the LAB file, rather than to do a single run using parameters with very wide
ranges in order to cater for all possibilities, and TRACS II allows this. The sets of shifts
generated by different runs may be merged later (Section 4.6).

4.5. Reduction of the shift set

The selection phase (Section 4.8) can sometimes be speeded up without loss of quality if the
number of shifts generated above can be reduced intelligently. The SIEVE process may be
applied separately to each or any of the sets of shifts generated above. It ranks each potential
shift using a combination of three attributes: an index which reflects its apparent efficiency (time
worked divided by time paid), the number of other potential shifts covering the piece of its work
which is covered by the smallest number of other shifts, and the average number of other
potential shifts covering the individual pieces of work making up the shift. The user specifies a
target number of shifts to be retained from the total set. The lowest ranked shifts are discarded,
and the last two attributes of the remaining shifts are updated; this process continues until the
target is reached. The user may then inspect a table showing for each of the apparent efficiency
indices the number of shifts discarded or retained, and may choose to reinstate a certain number
of these, in which case the most efficient are reinstated first.
SIEVE was very important some years ago because the ILP stage (Section 4.8) was rather
limited in the size of problem which it could handle. It has been used less often in recent
applications, given improvements in the sophistication of the generation process and in the
power of algorithms used to solve the ILP.

4.6. Merging sets of shifts

When more than one set of potential shifts has been generated by BUILD, these may be
merged at this stage to give a single set, with any duplication removed. This feature is frequently
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 445

used in UK bus applications involving mixed urban/rural operations in which rural-based

drivers commonly work to different agreements than do urban-based drivers.

4.7. Computation of penalties

The user may associate penalties with undesirable, albeit legal, features of shifts. These are
defined through the Workbench and applied to each remaining generated shift at this stage.
Users may repeat the process from this stage with different penalties in order to achieve desired
features of a schedule, but regular users normally work with a single set of penalties which they
have developed from experience.
Penalties are also used to generate ``minimum change'' schedules. These may be used either
when short-term changes are made to the vehicle schedule or when permanent small adjustments
are made, but it is desired for operational reasons to retain as much of the existing shift pattern
as possible. In these circumstances, TRACS II reads the existing driver schedule within the
BUILD process (Section 4.4), which ensures that shifts in that schedule, all of whose ROs still
exist, are present in the set generated and are assigned negative penalties.

4.8. Selection of a subset of shifts

The set covering model (1) is used to select a subset of the potential shifts which together
cover all the vehicle work. Since not all valid shifts have been generated, it may not be possible
to cover all pieces exactly once as would be required by a set partitioning approach. However
any objective to maximize schedule efficiency should implicitly aim to minimize the number of
pieces of work allocated to more than one shift. Any overlap may be edited afterwards, or may
indicate that a relaxation of parameters might be adopted in order to allow shorter shifts.
Given a problem with M pieces of work and N previously generated potential shifts, we can
define the set covering model to be:
X
N
Minimize Dj xj
jˆ1

X
N
Subject to Aij xj  1 for i ˆ 1, . . . , M
jˆ1

Plus any side constraints

xj ˆ 1 if shift j is used in the solution, 0 otherwise
Aij ˆ 1 if shift j covers piece of work i; 0 otherwise
Dj is determined by the objective function used …1†

Objective function. Three different objective functions are available in TRACS II. The user
may choose to minimize the number of shifts, minimize shift costs, or minimize a
lexicographically ordered combination of the two using a Sherali weighted cost function
described by Willers, Proll, and Wren (1995). Costs of individual shifts may be modified by
subjective penalties calculated during the previous stage. It is also possible to attach a cost to
any piece of work covered more than once.
446 A. WREN ET AL.

Constraints. There are necessary constraints on the model (1) which arise from the set
covering nature of the model. The user may also add side constraints governing the number of
shifts in the schedule in various combinations of shift type and depot, the total number of shifts
or the schedule cost.
Solution method. The integrality conditions are first relaxed to find the optimal linear
programming (LP) relaxation according to the objective function chosen, then a specialized
branch and bound process is adopted to obtain an integer solution representing a schedule. As
computing power and memory has become less of an issue, more shifts have usually been
generated than in the past, but storage space is not unlimited, and it is inefficient to store within
the LP process the large number of generated shifts which will not be used. For problems which
have a small generated set of shifts (currently fewer than 30,000) a dual approach, described by
Willers, Proll, and Wren (1995) is used. If the shift set is larger, then a primal SPRINT-like
(Bixby et al., 1992) technique described by Fores, Proll, and Wren (1999) is employed, which
ensures that all relevant generated shifts are considered when forming the optimal LP, but not
all are stored internally. Both primal and dual methods are accelerated by providing an initial
solution. Several heuristic methods of doing this are provided within TRACS II, and the default
constructs a schedule by sequentially looking at the currently uncovered piece of work having
the fewest shifts available to cover it. From these shifts, that covering the largest duration of
currently uncovered work is selected, thus attempting to cover all work with a small number of
shifts.
Both processes guarantee optimality of the relaxed model over all available shifts, but we do
not normally make all shifts available in the search for an integer solution. Smith and Wren
(1988) showed over a range of experiments that a good integer solution could normally be
obtained using only those ROs (basic ROs) used by shifts forming the basis of the relaxed LP,
and this conclusion has been confirmed by later practical cases. The user may therefore choose
to eliminate at this stage all shifts which use any non-basic ROs. If the primal approach has been
used, the process will have already selected a subset of shifts from all those available in such a
way as to retain a range of shifts which should be useful to the branch and bound search, as well
as to major iterations of the primal simplex method. Although adequate solutions are normally
found by proceeding to the branch and bound stage with this reduced subset, there will be other
potential shifts which are not present in this subset, but use only basic ROs. The user may
specify in advance that such shifts should be added to the subset, up to a current limit of 30,000
shifts in total.
If an objective function is used which requires shift minimization, a target cut is introduced
after the LP relaxation has been solved. The target cut is a side constraint which specifies that
the number of shifts must be exactly (or, optionally, at least) the rounded up number of shifts in
the LP solution. The branch and bound search requires a branching strategy, that is, rules to
select the node to explore next, and to construct the branches. There are a number of node
selection rules available in TRACS II which consider the extent of the integer infeasibility at the
node, and how much the objective function has degraded at this node. There are four entities
which can be branched on: shifts and relief opportunities, which proved satisfactory for single
depot problems, and types and depots, which have been added recently to aid the solution of
multi-depot problems (see Fores, Proll, and Wren, 2002). These are used in a hierarchical
manner; several such hierarchies are available. Within the branch and bound phase, dual
steepest edge (Forrest and Goldfarb, 1992) is used to solve the LP at each node. The branch and
bound process continues until either a specified number of nodes has been expanded or a
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 447

sequence of nodes has been encountered without any improvement to a previously found integer
solution.

4.9. Postprocessing
It is possible in the solution to a set covering problem that some pieces of work may be
included in more than one shift. The user is presented with a list of duplicate covers, and may
remove such overcover by manually deleting appropriate shifts. Alternatively, TRACS II will
remove overcover automatically, following heuristic rules, such as removing overcovered work
from the beginning or end of a shift, or in such a way as to reduce the work content of the longer
shifts. Such rules are acceptable to most users.

4.10. Output
The standard output from TRACS II is a list of shifts in summary form, showing for each
shift the signing-on and -off times, times and places where each vehicle is taken up and left,
durations of portions of work, and of meal breaks, the duration of the shift, and its cost. This is
appropriate for most bus companies, and can be read together with the vehicle schedule to give
full instructions to the drivers. An alternative form of output is similar to that used by most
British train operators. This incorporates information from the vehicle schedule to give full
details of each shift, including individual trips worked, and specification of journeys undertaken
as passengers. These output formats may be customized for individual users as required.

4.11. Shift validation

It is possible when a new user is being trained that the first schedule produced is less efficient
than expected. In such circumstances it may be necessary to examine carefully the parameters
being used in the LAB file, or to determine whether all travel opportunities have been correctly
set up in the TRA file. Often a valid shift is being prohibited for some reason such as wrongly set
parameters. The user may compare the existing manually produced schedule with that produced
by TRACS II. Where a good shift exists in the former, but not in the latter, the user should try
to work out why TRACS II has not used the shift. To this end, a CHECKER routine has been
provided. The user enters details of the shift in question, and CHECKER reports whether it
violates any rules. The user may then alter any appropriate parameters.
While an experienced user will normally obtain a good schedule at the first attempt, if there
has been a significant change in scheduling rules or in the nature of the operation it may be that
the resultant schedule is not as good as might have been hoped. Although in this case there is no
manual schedule for comparison, a good scheduler can often spot that shifts with some
characteristics are not present as expected. Again, CHECKER may be used to identify the
problem, so that parameters may be adjusted accordingly.

5. USER INTERFACE
The modules of TRACS II are run via the Workbench shown in Figure 2. This also provides an
interface for creating data files, setting parameters and penalties, cloning data sets to test
experimental ``what-if '' scenarios etc.
448 A. WREN ET AL.

Figure 2. The TRACS II workbench

The left-hand panel contains push buttons for each module in the suite, while the right-hand
panel contains information about the problem being tackled. Each module produces its own
report, either on screen, or as a log file, or both, in order to assist the user to rerun stages with
altered parameters as appropriate. The toolbar provides access to tools for the creation and
editing of data files. For example, the View/Edit button will produce a drop-down menu as in
Figure 3.
Vehicle work data in the TRA file is usually supplied to TRACS II from an external
timetabling system, and can be edited from the Workbench. Occasionally the data may be
entered manually. Labor rules may be prepared using a graphical user interface, part of which is
shown in Figure 4. This interface contains a total of seven screen pages for editing various
parameters that are applicable to most public transport applications. Users can specify different
types of shift required in the scheduling problem, for example for 4-day weeks, 5-day weeks,
part-time, etc. It also provides a facility to specify any extra payment to shifts with special
features.
The user need only provide details of the vehicle schedule, the labor agreement rules and any
penalties to be adopted. All other files are created as earlier stages of the solution process are
completed, but defaults used in these files may be altered by the user between stages.

6. SYSTEM DEVELOPMENT
TRACS II incorporates lessons learned through our use of the IMPACS driver scheduling
system which we first installed in London Transport Buses in 1984. Although this only allowed
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 449

Figure 3. View/Edit facilities in the TRACS II workbench

Figure 4. The labor agreement interface

450 A. WREN ET AL.

5000 potential shifts and a maximum of 50 nodes in the branch and bound process, it has
continued in use with few modifications until the present date. An enhanced version was
installed in Greater Manchester Transport in 1985, and subsequently adapted for PC use within
the BUSMAN system (Chamberlain and Wren, 1992), and installed in about thirty bus and light
rail companies until 1992. Work on TRACS II started in 1994; initially this was for rail driver
scheduling (Kwan et al., 1996), but throughout the development we ensured that it was capable
of solving a comprehensive range of bus driver problems accumulated during our earlier work.
Knowledge gained from new situations is continually incorporated into the system in order to
broaden the use of the application.
Large problems had to be decomposed for solution by IMPACS, and special routines were
developed both to set up subproblems and to carry over work from one subproblem to the next.
However, as computers became faster and as we introduced more powerful solution processes,
the need for decomposition decreased, and TRACS II can now solve problems with several
hundred shifts in the resultant schedule. The largest problem tackled to date in a single run had
about 2000 pieces of vehicle work and 1.5 million generated shifts.
Many column generation approaches (e.g., Desrochers et al., 1992; Caprara et al., 1999; Fahle
et al., 2002) construct new shifts incrementally as their need is recognized within the process,
using shortest path, dynamic programming, or constraint programming methods. We have
avoided this approach, as in our experience, the costs of individual shifts are not incremental
and cannot be determined until the whole shift is evaluated. Our approach also allows us to
separate the optimizing algorithms, which are the most complex part of the system, from the
problem specific domain, responsibility for which is largely delegated to the heavily
parameterized BUILD module. This is a major contribution to speedily achieving successes
with new clients.
Throughout the development and installation of TRACS II, we have worked closely with
users to enhance the system, improving the Workbench, adding to parameters and introducing
new recommended solution strategies. Solution strategies within the optimizing module are
controlled by two parameter files, default versions of which are created by the Workbench. The
principal parameters in the first of these define the objective function to be used and the length
of the branch and bound search. The second allows users more detailed control over the
optimizing algorithm, including choice of the shift subset to be submitted to the branch and
bound search, target cut, branching strategy and granularity of the search. The flexibility of this
approach allows users to choose strategies most appropriate to their needs and experiences, and
developers to readily perform the experimentation necessary to make recommendations on
solution strategies.
Although users have a wide range of parameters which govern solution strategies, most users
accept the default settings. However, the branch and bound process cannot guarantee that
solutions are found on every occasion, and adjustments may be made if particular difficulties
arise in obtaining solutions when new types of problem are presented. It is normal that once
good solutions have been obtained, the user can see the types of shift required. At that time it is
possible to adjust the shift generation parameters in the LAB file so that fewer potential shifts
are generated, after which the default parameters can normally be used to obtain a good
solution to the ILP.
The computer time required to obtain a good solution can vary, both in the BUILD process
and in the ILP, from a few minutes to many hours, but once suitable parameters have been set
for a new user it is normal to expect a solution to the largest problems in about an hour.
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 451

However, this is dependent on the user's own choice of shift generation parameters. A user who
can afford to wait several hours for a solution may choose very slack parameters in the hope
that this enables the ILP process to find a better solution, while if time is short, tight parameters
ensure that a reasonable solution is found quickly. In experimental work, we have allowed some
large problems to run for several days on a fast PC, but we rarely find any significant
improvement after a few hours. Such large problems are better solved by intelligently tightening
the generation parameters.
Some of the strategies are controlled by the process itself. For example, there is a range of
overall branching strategies in the optimization module which are controlled by one of the
parameter files. However, the default strategy and some others are themselves complex
strategies which instruct the process to adopt particular branching tactics depending on the
immediate situation.

7. EXPERIENCES
The two years, 1994±96, during which the original version of TRACS II was being developed,
coincided with the fragmentation of the national British Rail organization into twenty-five
separate operating companies, and preparation for their privatization. About half of these new
companies commissioned us to use TRACS II in order to evaluate new operating scenarios
involving either new labor rules, or new timetables, or both. For each of these companies, we
first ran TRACS II using their existing scenario, demonstrating every time that our system could
produce more efficient schedules than the existing manual ones; these first schedules were then
used as benchmarks against which the costs of the new scenarios could be evaluated. The largest
problem tackled during this period was for a regional rail operator with more than 400 daily
shifts and 20 different driver depots. It was necessary to decompose this into five overlapping
sub-problems, but some of the sub-problems were large and complex compared to our earlier
bus driver experiences, and led to our developing more powerful facilities in later program
versions. With the current version of TRACS II, decomposition is no longer necessary (Fores,
Proll, and Wren, 2001). The above exercises, and some subsequent projects, were carried out by
our in-house team on behalf of clients. The first application of the system in Reading (Wren and
Kwan, 1999), only 3 months after TRACS II had been ordered, resulted in an annual saving of
approximately £135,000, which was much more than the installation costs. Once this had been
achieved, management and unions together used the system to explore the effects of several new
operating rules proposed by the unions, resulting in the development of a new set of rules which
met some of the union's wishes at marginal increased operating costs.
In 1999 the UK's largest bus group, First, with a total fleet of about 10,000, set out to
evaluate a number of commercial scheduling packages. Meilton (2001) notes that the evaluation
exercise gave ``a comparison of the abilities of each system to solve an identical series of complex
duty scheduling problems and, as importantly, a meaningful comparison of the cost for each
acceptable answer.'' The outcome was that TRACS II was selected as the standard scheduling
tool for all twenty-five operating companies, each of which had its own set of scheduling rules,
varying from relatively simple broad rules, to very detailed and complex operations involving
several depots with different rules of interaction. First indications were that savings on wage
cost of more than 2% in situations where scheduling previously had been done manually, and
more than 0.5% where previously some form of computer-aided scheduling had been used were
452 A. WREN ET AL.

attainable. Meilton also notes that ``initially sceptical schedulers are persuaded on the merits of
the new system by both its ease of use and quality of results achieved.''
Implementation of TRACS II over the twenty-five operating companies of the First Group
(and 2 others) took place in less than 2 years from April 2000. During this period many new
features were added, both to meet needs of individual companies, and to enhance the TRACS II
system. This was accomplished by our working closely with a dedicated member of First staff
who helped to identify problems in advance, to help us set priorities, and who smoothed our
interaction with staff of the individual companies. The fact that so much was achieved in a
relatively short time for companies with wide variations in operating conditions and scheduling
rules testifies to our having achieved the objectives set out in Section 2. Training normally takes
two days, and schedulers have been able to immediately use the system to develop live schedules.
Within the same period our team has produced schedules for a major regional rail undertaking
and for a large intensive metropolitan railway. In all measurable cases, the schedules produced
have been recognized by the user as being significantly better than manual schedules in terms of
the objectives set. Some recent individual experiences follow.
TRACS II has been used to produce train driver schedules for a major train operating
company for their summer and winter schedules of 2001 and summer schedule of 2002. These
schedules are mainly for evaluation and for comparing to the manual schedules which were
produced as parallel exercises. The operation is mainly in the north of England and comprises
rural, inter-urban, and very intensive urban operations in major cities. One of the main features
is the detailed representation of ``route knowledge'' by TRACS II. In total, 63 portions of route
segment have been identified, some covering portions as small as 7 min driving time. In many
cases, drivers have to travel as passenger to move from one location to another. There are twelve
crew depots with the larger ones further divided, giving a total of 19 sub-depots. Some rural
depots must have a minimum number of shifts assigned to them. For the crew depots in major
cities, there is a preference to minimize the number of shifts assigned to them. The latter case is
achieved by penalizing potential shifts of undesirable depots. In all exercises, TRACS II
produced schedules with up to 8% fewer shifts than the manual solutions.
The work of a large urban bus depot, where drivers could work on any route, gave TRACS II
the opportunity to run on a problem in which there were in excess of 1274 ROs representing
nearly 2000 bus hours. Several schedules were produced detailing the work of 290 drivers.
Special techniques have been developed in TRACS II to solve successfully problems of this size.
They involve being able to identify those ROs which are most likely to be used in shifts in the
resulting schedule. In tackling schedules using the strategy of compiling a number of likely shifts
and then selecting those which cover all the bus work, the issue of how this should be applied to
large problems becomes critical. With the inappropriate selection of parameters it is easy to
compile literally millions of shifts. A more careful selection, while producing fewer shifts, might
overlook some relief opportunities which could be critical in producing the best schedule. By
being able to identify those relief opportunities most likely to be used by shifts in the final
schedule it is immediately possible to reduce the number of relief opportunites to be considered.
Parameters may be set at a level which, while impossible for the problem as a whole, will yield
shifts of the detail required to add significantly to the ease with which a good schedule can be
produced. These techniques were employed on this exercise and resulted in a manageable tool
for situations where a large network is to be scheduled all together.
TRACS II is in the course of evaluation by a major underground railway, and we have
recently undertaken comparisons with manual schedules on four of its operating lines, the
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 453

largest of which had about 250 daily shifts operating from three drivers' depots. In each case,
drivers had to travel as passengers on scheduled trains within their shifts, sometimes on services
of other companies. There were complex rules regarding movement of drivers between
platforms at some stations, with preference being given to changeovers taking place in stipulated
directions of travel. Drivers from some of the depots were allowed to sign on or off remotely
(i.e., without reporting to their home depot) at some agreed points and between certain hours.
The first two exercises were each spread over about a month, but the last two, using completely
new data and updated scheduling rules, were undertaken together within a period of 2 weeks. In
all cases, TRACS II produced schedules which were recognized by the company as being
workable and cheaper to operate than existing schedules.
It is worth commenting on the reaction of staff to the new scheduling system. During the
1970s when the use of computers in schedule construction was first discussed with bus operating
authorities, there were substantial misgivings as to whether the drivers' unions would accept
schedules that had been produced by computer. However, we do not know of any situation in
which this posed a real problem in practice. There have been several instances in which the
introduction of the computer system has enabled more driver-friendly schedule conditions to be
introduced at little or no extra cost, and in which unions and management have been able to use
the system together to develop new conditions acceptable to both. It is now generally accepted
by drivers' unions that if a schedule meets all the requirements negotiated with management,
then the way in which that schedule has been generated is irrelevant. Other staff directly affected
are of course the schedulers themselves, and almost all schedulers whom we have met, including
those who had no previous computer experience, have welcomed the introduction of the system.
It has removed drudgery from their lives, and has given them more time to experiment with
alternative operational practices.

8. CONCLUSIONS
Over a period of 35 years we have been developing driver scheduling methods in close associa-
tion with bus and train operating companies. Since 1984 our systems based on combinations of
heuristics and ILP have been in operational use in many locations. The TRACS II system
consistently produces schedules which are at least as good as those produced manually, and
normally considerably better. It is designed to be used on a PC, and has been proven to be
flexible, and easily adaptable to a wide range of operating scenarios. It is in regular use by a
wide range of bus companies, and has been applied by ourselves on behalf of rail companies,
both to produce operational schedules and to explore ``what if '' scenarios. The system has been
readily accepted by schedulers, even where there has been no previous computing experience.
The flexible approach which allows schedulers to influence the generation of potential shifts and
the solution strategy has been a major factor in achieving this.
While the system produces excellent schedules, we are continually learning, both from our
own experiences and from user feedback, how it may be improved. Many of the current and
imminent improvements are being made to details which are important to the user, but at a level
which cannot be described within this paper.
We have referred in Section 4.7 to the generation of ``minimum change'' schedules. Although
facilities for creating these have been requested by users, these have not yet been used in many
real situations, and we expect to improve the current facilities as we gain experience. At present
we are dealing with planned changes, that is, those which can be predicted several days or weeks
454 A. WREN ET AL.

in advance. Emergencies may of course arise in the course of operation, and although we are
looking at ways of treating these, the full power of TRACS II would be inappropriate. We can
learn from our scheduling experience how to develop quick and dirty approaches, but any
automatic approach has to rely on the drivers' acceptance of new workings which may extend
the working day and interfere with leisure activities. It is probable that any real-time system will
rely on the data within the TRACS II system, and will generate a number of possible quick
responses which may be discussed with the staff involved. We do not envisage any fully
automatic real-time system.

ACKNOWLEDGMENTS

We are grateful to many people within the bus and rail industries for their help over many years.
In particular, Michael Meilton of First has steered our implementations in that Group, and has
provided us with guidance and friendship. We are grateful to the UK Engineering and Physical
Sciences Research Council for substantial financial support, most recently under grants GR/
M23205, K79024 and K07256.

REFERENCES
Wren, A., ``Heuristics ancient and modern: Transport scheduling through the ages,'' J. Heuristics, 4, 87±100 (1998).
Fores, Sarah, Les Proll, and Anthony Wren, ``TRACS II: A hybrid IP/heuristic driver scheduling system for public
transport,'' J. Oper. Res. Soc., 53, 1093±1100 (2002).
Hamer, N. and L. Seguin, ``The HASTUS system: New algorithms and modules for the 90s,'' in M. Desrochers and
J.-M. Rousseau (eds), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1992, pp. 17±29.
Wolsey, L. A., Integer Programming, John Wiley & Sons, New York, 1998.
Johnson, C., E. L. Barnhart, G. L. Nemhauser, M. W. P. Savelsbergh, P. Vance, ``Branch-and-price: column generation
for solving huge integer programs,'' Oper. Res., 46, 316±329 (1998).
Kwan, R. S. K., A. S. K. Kwan, and A. Wren, ``Evolutionary driver scheduling with relief chains,'' Evol. Comput., 9,
445±460 (2001).
Smith, Barbara M., Colin J. Layfield, and Anthony Wren, ``A constraint programming pre-processor for a bus driver
scheduling system,'' in E. C. Freuder and R. J. Wallace (eds), Constraint Programming and Discrete Optimization,
American Mathematical Society, 2001, pp. 131±148.
Shen, Yingdong and Raymond S. K. Kwan, ``Tabu search for driver scheduling,'' in S. Voss and J. R. Daduna (eds),
Computer-Aided Scheduling of Public Transport, Springer-Verlag, Berlin, 2001, pp. 121±135.
Smith, B. M. and A. Wren, ``A bus crew scheduling system using a set covering formulation,'' Transporto. Res., 22A,
97±108, (1988).
Wren, A. and B. M. Smith, ``Experiences with a crew scheduling system based on set covering,'' in J. R. Daduna and
A. Wren (eds), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1988, pp. 104±118.
Parker, M. E., A. Wren, R. S. K. Kwan, ``Modelling the scheduling of train drivers,'' in J. R. Daduna,
I. Branco and J. M. P. Paixao (eds), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1995, pp. 213±235.
Willers, W. P., L. G. Proll, A. Wren, ``A dual strategy for solving the linear programming relaxation of a driver
scheduling system,'' Ann. Oper. Res., 58, 519±531 (1995).
Fores, Sarah and Les Proll, ``Driver scheduling by integer linear programmingÐthe TRACS II approach,'' in A. El Kamel
P Borne, M. Ksouri (eds.), Proceedings CESA'98 Computational Engineering in Systems Applications Symposium on
Industrial and Manufacturing Systems, LU Â nion des Chercheurs et Ingenieurs Scientifiques (UCIS), Villeneuve d'Ascq,
1998.
Fores, Sarah, Les Proll, Anthony Wren, ``An improved ILP system for driver scheduling,'' in N. H. M. Wilson (ed.),
Computer-Aided Transit Scheduling, Springer Verlag, Berlin, 1999, pp. 43±62.
Kwan, A. S. K., R. S. K. Kwan, M. E. Parker, A. Wren, ``Producing train driver schedules under different operating
strategies,'' in N.H.M. Wilson (ed.), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1999, pp. 129±154.
 A FLEXIBLE SYSTEM FOR SCHEDULING DRIVERS 455

Bixby, R. E., J. W. Gregory, I. L. Lustig, R. J. Marsten, and D. F. Shanno, ``Very large-scale linear programming: A case
study in combining interior point and simplec methods,'' Oper. Res., 40, 85±897, (1992).
Forrest, J. J. and D. Goldfarb, ``Steepest-edge simplex algorithms for linear programming,'' Math. Prog., 57, 341±374
(1992).
Chamberlain, M. and A. Wren, ``Developments and recent experience with the BUSMAN and BUSMAN II systems,''
in M. Desrochers and J.-M. Rousseau (eds), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1992,
pp. 1±15.
Kwan, A. S. K., R. S. K. Kwan, M. E. Parker, A. Wren, ``Producing train driver shifts by computer,'' in J. Allan,
C. A. Brebbia, R.J. Hill, G. Sciutto and S. Sone (eds), Computer in Railways V, Vol. 1: Railway Systems- and
Management, Computational Mechanics Publications, 1996, pp. 421±435.
Desrochers, M., J. Gilbert, M. Sauve, and F. Soumis, ``CREW-OPT: Subproblem modelling in a column generation
approach to urban crew scheduling,'' in M. Desrochers and J.-M. Rousseau (eds), Computer-Aided Transit
Scheduling, Springer-Verlag, Berlin, 1992, pp. 395±406.
Caprara, A., M. Fischetti, P. L. Guida, P. Toth, and D. Vigo, ``Solution of large-scale railway crew planning problems:
The Italian experience,'' in N.H.M. Wilson (ed.), Computer-Aided Transit Scheduling, Springer-Verlag, Berlin, 1999,
pp. 1±18.
Fahle, Torsten, Ulrich Junker, Stefan E. Karisch, Niklas Kohl, Meinolf Sellmann, and Bo Vaaben, ``Constraint
programming based column generation for crew assignment,'' J. Heuristics, 8, 59±82 (2002).
Fores, Sarah, Les Proll, and Anthony Wren, ``Experiences with a flexible driver scheduler,'' in S. Voss and J. R. Daduna
(eds), Computer-Aided Scheduling of Public Transport, Springer-Verlag, Berlin, 2001, pp. 137±152.
Wren, A., and R. S. K. Kwan, ``Installing an urban transport scheduling system,'' J. Sched., 2, 3±17 (1999).
Meilton, M., ``Selecting and implementing a computer aided scheduling system for a large bus company,'' in S. Voss and
J. R. Daduna (eds), Computer-Aided Scheduling of Public Transport, Springer-Verlag, Berlin, 2001, pp. 203±214.