Periodic Timetable Optimization in Public Transport

Christian Liebchen Institut f ur Mathematik, Kombinatorische Optimierung und Graphenalgorithmen, Technische Universit at Berlin, Germany liebchen@math.tu-berlin.de Summary. The timetable is the essence of the service oered by any provider of public transport. (Jonothan Tyler, CASPT 2006) Despite this observation, in the practice of planning public transportation, only some months ago OR decision support has still been limited to operations planning (vehicle scheduling, duty scheduling, crew rostering). We describe the optimization techniques that were employed in computing the very rst optimized timetable that went into daily service: the 2005 timetable of Berlin Underground. This timetables improved on both, the passenger travel times and the operating eciency of the company. The basic graph model, the Periodic Event Scheduling Problem (PESP), is known for 15 years and it had attracted many research groups. Nevertheless, we report on signicant progress that has been made only recently on issues like solution strategies or modeling capabilities. The latter even includes the integration of further planning tasks in public transport, such as line planning. On the theory side, we give a more precise notion of the asymptotical complexity of the PESP, by providing a MAXSNP-hardness proof as a kind of negative result. On the positive side, the design of more ecient algorithms gave rise to a much deeper understanding of cycle bases of graphs, another very hot topic in discrete mathematics during the last three years. In 2005, this culminated in both, drawing the complete map for the seven relevant classes of cycle bases, and the design of the fastest algorithms for the Minimum Directed Cycle Basis Problem and for the Minimum 2-Basis Problem. The book version of this extended abstract is available as reference [8].

1 Timetabling
It is a very important competitive advantage of public transport to be much less expensive than a taxi service. This requires many passengers to share the same vehicle. Typically, this is achieved by oering public transport along xed sets of routes, the lines. These serve as input to timetabling. This work has been supported by the DFG Research Center Matheon in Berlin.

30

Christian Liebchen

There is a large toolbox of dierent types of timetables, which we introduce from the most general one to the most specialized one: timetables that are composed of individual trips, periodic timetables, i.e. the headway between any two successive trips of the same line is the same, symmetric periodic timetables, and so-called Integrated Fixed-Interval Timetables. Here, a periodic timetable is called symmetric, if for every passenger the transfer times that he faces during his outbound trip are identical to the transfer times during his return trip, which here is assumed to have the same route. In particular, the periodic timetables of most European national railway companies are indeed symmetric, because marketing departments consider this being a competitive advantageat least in long-distance trac. Theorem 1 ([7]). There exist example networks showing that each more specialized family of timetables causes a nominal loss in quantiable criteria, such as average passenger waiting time. We are only aware of periodic timetables being able to clearly outweigh their nominal loss (when comparing with general irregular timetables) by adding benet in qualitative criteria. Hence, in the remainder we focus on periodic timetables. Typically, the period time T varies over the day. For instance, Berlin Underground distinguishes rush hour service (T = 4 minutes), normal service (T = 5 minutes), weak trac service (T = 10 minutes, when retail shops are closed), and night service (T = 15 minutes, only on weekends). Computing the timetable thus decomposes into computing a periodic timetable for each period time, and nally glue these together.

2 A Model for Periodic Timetabling

A literature review of dierent models for periodic scheduling reveals that the most promising earlier studies on medium-sized networks are based on the Periodic Event Scheduling Problem (Pesp, [18]), see [17, 14, 9, 16]. The vertices in this graph model represent events, where an event v V is either an arrival or a departure of a directed line in a specic station. A timetable assigns to each vertex v a point in time v [0,T ) within the period time T . Constraints may then be given in the following form.

T-Periodic Event Scheduling Problem (T-Pesp)

Instance: Task: A directed graph D = (V,A) and vectors ,u QA . Find a vector [0,T )V that fullls ( v u
a)

mod T ua

(1)

(or v u [ a ,ua ]T , for short) for every arc a = (u,v ) A, or decide that none exists.

Periodic Timetable Optimization in Public Transport

31

In Figure 1 we provide an example instance of T -Pesp, which contains the events of two pairs of directed lines and two stations.

Ol
Osloer Strae
[1.5,6.5; wa ]T

[0.5,1.5; wa ]T

[7.0,7.0]T

[1.5,1.5]T

Fig. 1. A Pesp model for two lines and two stations Here, the straight arcs model either stops (within the black box that represents the station) or trips, and the dotted arcs model either passenger transfers or turnarounds of the trains. Besides these most elementary requirements, there have been modeled most practical requirements that railway engineers have ([11]). This even includes decisions of line planning and vehicle scheduling, which traditionally were treated as fully separate planning steps ([1]). Unfortunately, this modeling power has its price in terms of complexity. Theorem 2. Let a set of PESP constraints be given. Finding a timetable vector that satises a maximum number of constraints is MAXSNP-hard. To make T -Pesp accessible for integer programming (IP) techniques, the modulooperator is resolved by introducing integer variables: min wT (B T + T p) s.t. B T + T p u BT + T p [0,T )V p {0,1,2}A .

(2)

Here, the matrix B is the vertex-arc incidence matrix of the constraint graph D = (V,A). But this is not the only way to formulate T -Pesp as an IP. Rather, we may replace the vertex variables (or node potentials) which carry time information with arc variables x (tensions), and/or replace the integer arc variables p with integer cycle variables z . This way, we end with the following integer program ([14]) min wT x s.t. x u x Tx Tz = 0 x QA z ZB ,

(3)

32

Christian Liebchen

where denotes the arc-cycle incidence matrix of an integral cycle basis B of the constraint graph D = (V,A). There have also been identied lower and upper bounds on these integer variables z . Theorem 3 ([15]). Let C be an oriented circuit and zC the integer variable that we associate with it. The following inequalities are valid 1 1 ua =: z C zC z C := ua a a . T T a C + a C a C + a C The following rule-of-thumb could be derived from empirical studies. Remark 1 ([4]). The shorter a circuit C B with respect to the sum of the spans ua a of its arcs, the less integer values the corresponding variable zC may take. Moreover, the less values all the integer variables may take, the shorter the solution times for solving this IP.

3 Integral Cycle Bases

In a directed graph D = (V,A), we consider oriented circuits. These consist of forward arcs and maybe also backward arcs, such that re-orienting the backward arcs yields a directed circuit. The incidence vector C {1,0,1}A of an oriented circuit C has a plus (minus) one entry precisely for the forward (backward) arcs of C . Then, the cycle space C (D) can be dened as C (D) := span({C | C is an oriented circuit of D}). A cycle basis B of C (D) is a set of oriented circuits, which is a basis of C (D). An integral cycle basis allows to combine every oriented circuit of D as an integer linear combination of the basic circuits. Fortunately, in order to decide upon the integrality of a cycle basis, we do not have to check all these linear combinations. Lemma 1 ([4]). Let be the arc-cycle incidence matrix of a cycle basis. For two submatrices 1 ,2 with rank(1 ) = rank(2 ) = rank( ), there holds det 1 = det 2 . (4)

Denition 1 (Determinant of a cycle basis, [4]). Let B be a cycle basis and 1 as in the above lemma. We dene the determinant of B as det B := |det 1 |. Theorem 4 ([4]). A cycle basis B is integral, if and only if det B = 1. According to Remark 1, in the application of periodic timetabling we seek for a minimum integral cycle basis of D. To illustrate the benet of short integral cycle we provide the following example. (5)

Periodic Timetable Optimization in Public Transport

33

Fig. 2. The sunower graph SF(3), and a spanning tree

Example 1. Consider the sunower graph SF(3) in Figure 2. Assume each arc models a Pesp constraint of the form [7,13]10 , i.e. subject to a period time of T = 10. According to the initial IP formulation (2), we could deduce that in order to identify an optimum timetable simply by considering every possible timetable vector we need to check |{0, . . . ,9}||V | = 1,000,000 vectors . Alternatively, we might check for |{0,1,2}||A| = 19,683 vectors p to pervade the search space. It will turn out, that these two perspectives have much redundancies within them. In contrast, the valid inequalities of Theorem 3 reveal that every 4-circuit C in D yields an integer variable zC which may only take the three values {1,0,1}. Even better, a triangle C in D induces a xed variable zC . Thus, the integral cycle basis B that can be derived from the spanning tree F (Fig. 2 on the right) already reduces the upper bound on the size of the search space to only 1 1 1 3 = 3 possible vectors for z . Moreover, considering the minimum cycle basis of D which for this graph turns out to be integral as it consists of the circuits that bound the four nite faces of this plane graph we end with just one single vector z describing the complete instance. Ideally, we would like to compute a minimum integral cycle basis of D according to the edge weights ua a . Unfortunately, we are not aware of the asymptotical complexity of this combinatorial optimization problem. However, recently there has been achieved much progress on the the asymptotical complexity for the corresponding minimum cycle basis problems for related classes of cycle bases, see [13, 2] and referenced therein. We depict these results in Figure 3. Notice that any of these classes demands for specic algorithms, because none of these problems coincide ([13]).

4 Summary of Computational Results

Earlier, in several autonomous computational studies, there had been applied various algorithms to periodic timetabling. We have executed the rst unied computational study, which covers algorithms as variegated as a Cut-and-Branch Algorithm for Integer Programs, Constraint Programming, and even Genetic Algorithms ([12]). To any of these, quite a number of dierent promising parameter settings was applied. In particular for Integer Programming this amounts to hundreds of dierent runs on ve dierent data sets. All the data sets have been provided to us by industrial partners, and they range from long-distance trac over regional trac down to undergrounds, comprising between 10 and 40 commercial lines each. With respect to both solution quality and independence of parameter settings, both our Genetic Algorithm (GA) and our Cut-and-Branch Algorithmwhich is

34

Christian Liebchen

strictly TUM weakly integral undirected directed

O(n) NPC (open) (open) (open) O (m2 n . . . ) O (m3 n . . . )

2-bases

Fig. 3. Map of the complexity of the seven variants of the Minimum Cycle Basis Problem for general graphs ([4, 13])

using CPLEX 9.1perform considerably well. On the one hand, IP techniques turn out to be extremely sensitive with respect to the choice of several important parameters. On the other hand, in particular for medium-sized instances for which IP techniques still attain an optimum solution, the quality achieved by the GA is somewhat worse.

5 Improvements for Berlin Underground

Most important, in a long-term cooperation with Berlin Underground we continuously kept on improving our mathematical models of the real world ([10, 6]). Finally, in 2004 we were able to formulate a mathematical program which covered all the practical requirements that the practitioners have. As a consequence, the optimum solution that was computed by our algorithms convinced Berlin Underground: By December 12, 2004, our timetable became the rst optimized timetable that went into servicepresumably worldwide. This may be compared to the fact that only in operations planning (vehicle scheduling, duty scheduling), Operations Research had already entered the practice. Compared to the former timetable, with our timetable the passengers of Berlin Underground are oered simultaneously improvements in two key criteria, which typically are conicting: transfer waiting time and dwell time of trains. In more detail, our achievements are: The number of transfers, for which a maximum transfer waiting time of 5 minutes can be guaranteed, increases from 95 to 103 (+8%). The maximum dwell time of any train in the network was reduced from 3.5 minutes to only 2.5 minutes (30%). The timetable could even be operated with one train less.

Periodic Timetable Optimization in Public Transport

35

The part of the network in which the most signicant improvements have been achieved is given in Figure 4. Network waiting time charts emerged as a by-product from our cooperation with Berlin Underground ([6, 7]). Such charts constitute the rst visualization of the transfer quality of a timetable. In particular, they made the discussion of pros and cons of dierent timetables most ecient, as for instance long transfer waiting times (marked in black) along important transfers (marked as bold arcs) become obvious.

04.5

Sn Fp
02.0 08.5

Sn

Fp Be

06.0

Be

Fig. 4. Network waiting time charts for an excerpt of the Berlin subway network before and after invoking mathematical optimization The successful transfer from theory to practice has even been reected by articles and interviews in nationwide newspapers and radio transmissions: Berliner Zeitung, November 9, 2005, in German ([3])
http://www.berlinonline.de/berliner-zeitung/archiv/.bin/dump.fcgi/2005/1109/wissenschaft/0002/index.html

Deutschlandfunk, December 9, 2005, 16:35h, in German

http://www.dradio.de/dlf/sendungen/forschak/446751/

References
1. Michael R. Bussieck, Thomas Winter, and Uwe Zimmermann. Discrete optimization in public rail transport. Mathematical Programming B, 79:415444, 1997. 2. Ramesh Hariharan, Telikepalli Kavitha, and Kurt Mehlhorn. A Faster Deterministic Algorithm for Minimum Cycle Bases in Directed Graphs. In Michele Bugliesi et al., editors, ICALP, volume 4051 of Lecture Notes in Computer Science, pages 250261. Springer, 2006. 3. Reinhard Huschke. Schneller Umsteigen. Berliner Zeitung, 61(262):12, 2005. Wednesday, November 9, 2005, In German. 4. Christian Liebchen. Finding short integral cycle bases for cyclic timetabling. In Giuseppe Di Battista and Uri Zwick, editors, ESA, volume 2832 of Lecture Notes in Computer Science, pages 715726. Springer, 2003. 5. Christian Liebchen. A cut-based heuristic to produce almost feasible periodic railway timetables. In Sotiris E. Nikoletseas, editor, WEA, volume 3503 of Lecture Notes in Computer Science, pages 354366. Springer, 2005.

36

Christian Liebchen 6. Christian Liebchen. Der Berliner U-Bahn Fahrplan 2005 Realisierung eines mathematisch optimierten Angebotskonzeptes. In HEUREKA 05: Optimierung in Transport und Verkehr, Tagungsbericht, number 002/81. FGSV Verlag, 2005. In German. 7. Christian Liebchen. Fahrplanoptimierung im PersonenverkehrMuss es immer ITF sein? Eisenbahntechnische Rundschau, 54(11):689702, 2005. In German. 8. Christian Liebchen. Periodic Timetable Optimization in Public Transport. dissertation.de, 2006. PhD thesis. 9. Thomas Lindner. Train Schedule Optimization in Public Rail Transport. Ph.D. thesis, Technische Universit at Braunschweig, 2000. 10. Christian Liebchen and Rolf H. M ohring. A case study in periodic timetabling. Electr. Notes in Theoretical Computer Science, 66(6), 2002. 11. Christian Liebchen and Rolf H. M ohring. The modeling power of the periodic event scheduling problem: Railway timetables and beyond. Preprint 020/2004, TU Berlin, Mathematical Institute, 2004. To appear in Springer LNCS Volume Algorithmic Methods for Railway Optimization. 12. Christian Liebchen, Mark Proksch, and Frank H. Wagner. Performance of algorithms for periodic timetable optimization. To appear in Springer LNEMS PProceedings of the Ninth International Workshop on ComputerAided Scheduling of Public Transport (CASPT). To appear. 13. Christian Liebchen and Romeo Rizzi. Cycles bases of graphs. Technical Report 2005-018, TU Berlin, Mathematical Institute, 2005. accepted for publication in Discrete Applied Mathematics. 14. Karl Nachtigall. Periodic Network Optimization and Fixed Interval Timetables. Habilitation thesis, Universit at Hildesheim, 1998. 15. Michiel A. Odijk. A constraint generation algorithm for the construction of periodic railway timetables. Transp. Res. B, 30(6):455464, 1996. 16. Leon W.P. Peeters. Cyclic Railway Timetable Optimization. Ph.D. thesis, Erasmus Universiteit Rotterdam, 2003. 17. Alexander Schrijver and Adri G. Steenbeek. Dienstregelingontwikkeling voor Railned. Rapport CADANS 1.0, Centrum voor Wiskunde en Informatica, December 1994. In Dutch. 18. Paolo Serani and Walter Ukovich. A mathematical model for periodic scheduling problems. SIAM Journal on Discrete Mathematics, 2(4):550581, 1989.