Einführung in Verkehr und Logistik [1.15ex] (Bachelor) [1.15ex
Transcription
Einführung in Verkehr und Logistik [1.15ex] (Bachelor) [1.15ex
WS 13/14 Einführung in Verkehr und Logistik 1 / 56 Einführung in Verkehr und Logistik (Bachelor) Transport Demand Univ.-Prof. Dr. Knut Haase Institut für Verkehrswirtschaft Wintersemester 2013/2014, Dienstag 10:15-11:45 Uhr, Phil E WS 13/14 Einführung in Verkehr und Logistik Service Production Process in Public Transport 1. Demand Model 2. Infrastructure (e.g. tracks and stations) 3. Tariff zone planning 4. Line planning (routes and frequencies) 5. Timetabling 6. Vehicle scheduling 7. Crew scheduling 8. Rostering 9. Dispatching (vehicles and staff) I I Simultaneous planning approaches, e.g. simultaneous vehicle and crew scheduling, provide significant cost reduction potentials.1 Demand effects should be integrated in line planning and timetabling.2 1 See 2 See [HDD01]. [KH08]. 3 / 56 WS 13/14 Einführung in Verkehr und Logistik 5 / 56 Transport Demand What are discrete choices? I I I I I I Course of study Transport mode Port choice by a shipowner Beer brand Shopping facility ... Discrete choice appear in every day of life! WS 13/14 Einführung in Verkehr und Logistik An example The choice between taking the bus or car might be influenced by I I I I I I I I Cost Travel-time Income Flexibility Interchanges Security Eco friendliness ... 6 / 56 WS 13/14 Einführung in Verkehr und Logistik 7 / 56 How to get the data and what to do with it? Empirical analysis I Sample; we distinguish in general I revealed preferences (chosen alternative) I stated preferences (stated alternative) I A sample is used to I I I I I Specification of utility functions Demand forecast Analysis of willingness to pay Evaluation of infrastructural improvements Service design WS 13/14 Einführung in Verkehr und Logistik Numerical example Transport mode choice of forwarders in a region: Would you choose combined cargo if the time for intial leg and final leg is I 1 hour? I 2 hours? I 3 hours? 8 / 56 WS 13/14 Einführung in Verkehr und Logistik 9 / 56 Contingency table Combined cargo yes (i = 1) no (i = 2) time for initial and final leg 1h (k = 1) 2h (k = 2) 3h (k = 3) 100 100 50 200 300 250 300 400 300 250 750 1000 p̂(i = 1) = 250=1000 = 0:25 p̂(i = 1; k = 2) = 100=1000 = 0:10 p̂(i = 1) = 3 P p(i = 1; k) = 100 1000 p̂(k = 3ji = 1) = 50=250 = 0:2 k=1 WS 13/14 + 100 1000 + 50 1000 = 0:25 Einführung in Verkehr und Logistik p̂(i = 1; k = 2) = p̂(i = 1)p̂(k = 2ji = 1) = 0:25 0:40 = 0:10 = p̂(k = 2)p̂(i = 1jk = 2) = 0:40 0:25 = 0:10 Let p̂(i) > 0 and p̂(k) > 0: p̂(k ji) p̂(i jk) = p̂(i ; k)=p(i) = p̂(i ; k)=p(k) Assumption: p̂(i jk) is constant over all periods p̂(k) might change over periods! 10 / 56 WS 13/14 Einführung in Verkehr und Logistik 11 / 56 A behavioural model p(i = 1jk = 1) 1 p(i = 1jk = 2) = 2 p(i = 1jk = 3) = 3 3 Parameter p̂(i = 2jk) = 1 y=n) p̂(i = 1jk) = !not an additional parameter (binary case: Estimate a simple model ˆ1 = 100=300 = 0:333 ˆ2 = 100=400 = 0:250 ˆ3 = 50=300 = 0:167 Hypothesis: The smaller the time for initial leg and final leg the larger the choice probability of combined cargo WS 13/14 Einführung in Verkehr und Logistik 12 / 56 Standardfehler der Schätzung I I I I I I I I Bernoulli-distributed variable (urn experiment) Sum of Bernoulli-distributed variable is binomial distributed. p Standard error of the mean of a random variable: 2 =N 2 : Variance N : Sample size Variance of a Bernoulli-distributed variable: (1 ) : Probability of occurrence p ˆk (1 ˆk )=Nk Estimator for standard errors: ŝk = Approximation for the 95% confidence intervall (c.i.): [ˆ k ˆk =ŝk t-statistic : t-value > 2 !significant different from zero 2 ŝk ] WS 13/14 Einführung in Verkehr und Logistik 13 / 56 Results for our small example ˆk k =1 k =2 k =3 0.333 0.250 0.167 WS 13/14 ŝk 0.027 0.022 0.022 95% c.i. [0.279,0.388] [0.207,0,293] [0.124,0.210] t-value 12.247 11.547 7.746 Einführung in Verkehr und Logistik Apply our model Assume we obtain the following data for a region (potentials for combined cargo): time of initial and final leg 1h 2h 3h Market share: 22.73% transport weight in tons per year total combined cargo (estimates) 1,200,000 400,000 2,400,000 600,000 3,000,000 500,000 6,600,000 1,500,000 14 / 56 WS 13/14 Einführung in Verkehr und Logistik 15 / 56 Construction of additional terminals time of inital and final leg 1h 2h 3h transport weight in tons per year total combined cargo (estimates) 1,500,000 500,000 2,700,000 675,000 2,400,000 400,000 6,600,000 1,775,000 !Market share increases up to 29% WS 13/14 Einführung in Verkehr und Logistik Properties of estimators I Unbiased I Efficient I For non-linear models we have only asymptotical properties: - consistent: N ! 1 estimator = true value - asymptotically efficient: there is no consistent estimator with a small standard error I Discrete choice models are non-linear models! 16 / 56 WS 13/14 Einführung in Verkehr und Logistik 18 / 56 Assumptions I I I Individuals (n = 1; : : : ; N) choose between only two alternatives (i = 1; 2) Individual (or observation) n chooses alternative i with largest utility Assume a linear utility function Uni that consists of I deterministic (representative) utility Vni and I stochastic utility ni , such that I Uni = Vni + ni I ni is independent and identically extreme value distributed (IID EV) with WS 13/14 f (ni ) = e ni e F (ni ) = e e ni e ni Einführung in Verkehr und Logistik 19 / 56 Density function and cumulative density function of EV 1 F (!ni ) = e−e !ni 0,75 0,5 0,25 f (!ni ) = e−!ni e−e !ni ! 5 4,5 4 3,5 3 2,5 2 1,5 1 0,5 0 -0,5 -1 -1,5 -2 0 WS 13/14 Einführung in Verkehr und Logistik 20 / 56 Central assumption The difference of two independently extreme value distributed variables n = ni 0 ni , is logistically distributed: f (n ) = F (n ) = e n (1 + e n )2 1 1 + e n >0 is a scale parameter WS 13/14 Einführung in Verkehr und Logistik 21 / 56 Choice probabilities I n chooses i = 1, iff Un1 > Un2 Since Uni is random (due to ni ) we can only compute the propability of Un1 > Un2 Pn (i = 1) = P(Un1 > Un2 ) = P(Vn1 + n1 > Vn2 + n2 ) = P(n1 = P(n2 n2 > Vn2 n1 < Vn1 = P(n < Vn1 Vn2 ) Vn1 ) Vn2 ) WS 13/14 Einführung in Verkehr und Logistik 22 / 56 Choice probabilities II Pn (i = 1) = P(n2 n1 < Vn1 = P(n < Vn1 = F (Vn1 = = Vn2 ) Vn2 ) Vn2 ) 1 1 + e (Vn1 Vn2 ) e Vn1 e Vn1 + e Vn2 Binary Logit Model e Vn1 Pn (i = 1) = Vn1 e + e Vn2 is not identified and has to be fixed to an arbitrary value (1 e.g.) WS 13/14 Einführung in Verkehr und Logistik Normalisation We can not measure the absolute value of utility. A multiplication of utility of each alternative with a positive constant does not change choice probabilities (and thus choice process). !Normalisation of = 1 !Variance of ni : 2 =6 23 / 56 WS 13/14 Einführung in Verkehr und Logistik 24 / 56 Reference alternative An addition of a constant to the utility of each alternative does not influence the choice probabilities !Constrain coefficients of the refernce alternative to a fixed value (e.g. 0) Un1 = 1 + 3 TC1 + 4 TT1 + 6 INCn + n1 (1) Un2 = 2 + 3 TC2 + 5 TT2 + 7 INCn + n2 (2) 2 = 7 = 0 WS 13/14 Einführung in Verkehr und Logistik 25 / 56 Choice probability of alternative 1 P(i = 1) Pn (i = 1) 1 0,75 0,5 Pn (i = 1) = 1 1 + e−Vn1 Vn2 = 0 Vn1 5 4 3 2 1 0 -1 -2 -3 -4 0 -5 0,25 WS 13/14 Einführung in Verkehr und Logistik 27 / 56 Multinomial Logit Model (MNL) Assumptions I sames as binary logit I ni is iid EV over all i WS 13/14 and n f (ni ) = e ni e F (ni ) = e e ni e ni Einführung in Verkehr und Logistik MNL choice probabilities3 Pn (i) = P[ 8j = 6 i : Uni > Unj ] = P[ 8j = 6 i : Vni + ni > Vnj + nj ] = P[ 8j = 6 i : nj Z = I ( 8j = 6 i : nj Vni Vnj ] ni Vni Vnj )f (n )dn ni if ni iid EV = e Vni P V nj j e Comments: I I () Indicator function I We may solve the integral even if ni is not iid EV ! Nested-, Cross-Nested, Mixed Logit 3 See [Tra03, pp. 38-41] 28 / 56 WS 13/14 Einführung in Verkehr und Logistik 29 / 56 Some comments I We may write MNL choice probabilities as Pn (i) = P e Vni Vnj j 2Cn e Scale parameter Cn Choice set of individual n =0 !1 j Cn j= 2 WS 13/14 ) even distributed choice probabilities ) deterministic choices ) binary logit Einführung in Verkehr und Logistik Independence from irrelevant alternatives (IIA)4 I Stems directly from the ”independent” part of the iid EV assumption on ni I Ratio of choice probabilities of two alternatives is independent from the presence or absence of another alternative (or their attributes) I Might be inadequat some times. If so, then make other assumption about distribution of ni I !Nested-Logit, Mixed Logit etc. Proof IIA Pn (i) Pn (i 0 ) = e Vni 0 e Vni P = Vnj Vnj j e j e P e Vni e Vni 0 = e Vni = 4 See [Tra03, pp. 49-54] Vni 0 30 / 56 WS 13/14 Einführung in Verkehr und Logistik 31 / 56 Red bus - blue bus paradoxon5 Example Pn (’Car’) = 1 1 ; Pn (’Red Bus’) = 2 2 Now add a blue bus to Cn Pn (’Car’) = 1 1 1 ; Pn (’Red Bus’) = ; Pn (’Blue Bus’) = 3 3 3 We would expect Pn (’Car’) = 5 See 1 1 1 ; Pn (’Red Bus’) = ; Pn (’Blue Bus’) = 2 4 4 [BAL85, pp. 51-55] WS 13/14 Einführung in Verkehr und Logistik IIA holds strictly only on individual level Example: choice of shopping facility Group with car no car frequency CBD 20 % 80 % 50 % Outlet park 1 80 % 20 % 50 % Add a new alternative Outlet park 2 Group with car no car frequency CBD 11.11 % 66.66 % 38.88 % Outlet park 1 44.44 % 16.66 % 30.56 % Outlet park 2 44.44 % 16.66 % 30.56 % ) The more socio-economic variables we use the less critical is IIA on ”market-share”-level 32 / 56 WS 13/14 Einführung in Verkehr und Logistik 33 / 56 Specification and identification Un1 = 1 + 4 TC1 + 4 TT1 + 6 INCn + n1 Un2 = 2 + 4 TC2 + 5 TT2 + 7 INCn + n2 Un3 = 3 + 4 TC3 + 6 TT3 + 8 INCn +n3 | {z } Vn3 I 1 , 2 , 3 : alternative-specific constants (ASCs) I Only jCn j 1 ASCs are identified I ASCs reflect the expectation value of ni only if a full set of ASCs (jCn j 1) is employed. I Alternative-specific specifications of socio-economic variables (depend on n only!) - like income - need normalisation, too. WS 13/14 Einführung in Verkehr und Logistik Specification of utility with categorial variables (k) Attribute (variable) Value (l ), reference category (1) Age (2) Income (3) Gender teen (t), adult (a), old (o) high (h), low (l ) female (f ), male (m) Zkl : Dummy variable of attribute k with value l i0 + it Z1;t + ia Z1;a + ih Z2;h + if Z3;f Segment of population (old, low income, male): 0 ASC for segment (adult, high income, male): 0 + a + h 34 / 56 WS 13/14 Einführung in Verkehr und Logistik 36 / 56 Estimation of ’s by Maximum Likelihood6 How do we get the ’s? Definition yni =1, if n has chosen i (0, otherwise) Re-write logit choice probabilities as J Y Pn (i) = (Pn (i))yni i =1 Since ni are iid over n as well we can write the likelihood function as L( ) = N J Y Y (Pn (i))yni n=1 i =1 Note, is the coefficient vector of interest! That is, all ik to be estimated. 6 See [Tra03, pp. 64-67] WS 13/14 Einführung in Verkehr und Logistik 37 / 56 The log-Likelihood makes estimation easier and obtains the same extreme points as L( ) ! N J N J Y Y X X yni L( ) = ln L( ) = ln (Pn (i)) = yni ln Pn (i) n=1 i =1 Substitute Pn (i) with Pe L( ) = N X J X Vni V e nj j = = n=1 i =1 ! e Vnj N J X X yni Vni n=1 i =1 N J X X Vni j n=1 i =1 N J X X ! e P yni ln n=1 i =1 yni ln n=1 i =1 yni X k ik znik ! X ! e Vnj j N J X X n=1 i =1 yni ln X j P e k jk znjk ! WS 13/14 Einführung in Verkehr und Logistik Maximise 38 / 56 L( ) Start ˆ = 0 : L(ˆ) = 0 N ln J (if Cn is the same for all n) ˆ Vector of estimated coefficients that maximises L( ) Goodness-of-fit Pseudo-R 2 : LR = 1 l (ˆ )=l (0) Software for estimation I I I I BIOGEME LIMDEP, NLOGIT SPSS R WS 13/14 Einführung in Verkehr und Logistik BIerlaire Optimization toolbox for GEv Model Estimation7 MNL is a special case of a Generalised Extreme Value (GEV) Model (nowadays: Multivariate Extreme Value (MEV)) http://biogeme.epfl.ch public domain, cross plattform Invoke via console (f.e. DOS prompt: cmd.exe): biogeme mymodel mysample.dat I We write the model to mymodel.mod I .mod file may be manipulated by any texteditor (notepad++ is advised) I Data for estimation has to be stored in mysample.dat I Results can be found in output mymodel.rep 7 See [Bie08]. 40 / 56 WS 13/14 Einführung in Verkehr und Logistik 41 / 56 Example A choice problem Students (n) with no car available chose on their commute to university either to walk (i = 1) or go by bike (i = 2) or to take the bus (i = 3). The choice depends on travel-time in minutes dni and gender sn with female students indicated by 1. Now we write our deterministic utilities Vn1 Vn2 V3n 0;1 + 1;1 d1;n + 2;1 sn = 0;2 + 1;2 d2;n + 2;2 sn = 0;3 + 1;3 d3;n + 2;3 sn = For identification purposes: 0;1 = 1;1 = 2;1 = 0 That is, Vn1 = 0 WS 13/14 Einführung in Verkehr und Logistik BIOGEME syntax mymodel.mod [Choice] Choice [Beta] // Name b01 b02 b03 b11 b12 b13 b21 b22 b23 Value 0 0 0 0 0 0 0 0 0 LowerBound -10000 -10000 -10000 -10000 -10000 -10000 -10000 -10000 -10000 UpperBound 10000 10000 10000 10000 10000 10000 10000 10000 10000 status (0=variable, 1=fixed) 1 0 0 1 0 0 1 0 0 [Utilities] // Id Name Avail linear-in-parameter expression 1 Foot av1 b01 * one + b11 * d1 + b21 * s 2 Bike av2 b02 * one + b12 * d2 + b22 * s 3 Bus av3 b03 * one + b13 * d3 + b23 * s [Expressions] one = 1 av1 = 1 av2 = 1 av3 = 1 [Model] $MNL 42 / 56 WS 13/14 Einführung in Verkehr und Logistik 43 / 56 Data mysample.dat ID Choice 1 1 2 1 3 1 4 2 5 1 : : 95 3 96 1 97 1 98 2 99 1 100 1 WS 13/14 d1 6 8 12 21 6 : 51 18 6 33 19 15 d2 8 9 9 12 7 : 21 11 9 15 11 12 d3 13 13 15 17 12 : 24 14 12 16 13 16 s 0 1 1 0 0 : 0 1 0 1 0 1 Einführung in Verkehr und Logistik Results: model statistics mymodel.rep Model: Number of estimated parameters: Number of observations: Number of individuals: Null log-likelihood: Cte log-likelihood: Init log-likelihood: Final log-likelihood: Likelihood ratio test: Rho-square: Adjusted rho-square: Final gradient norm: Diagnostic: Iterations: Run time: Multinomial Logit 6 100 100 -109.861 -88.991 -109.861 -68.002 83.718 0.381 0.326 +4.082e-07 Convergence reached... 11 00:00 44 / 56 WS 13/14 Einführung in Verkehr und Logistik 45 / 56 Results: coefficient estimates Utility parameters ****************** Name Value Std err t-test p-val Rob. std err Rob. t-test Rob. p-val ---- ----- ------- ------ ----- ------------ ----------- ---------b01 0.00 --fixed-b02 -5.57 1.15 -4.84 0.00 1.01 -5.50 0.00 b03 -10.4 2.17 -4.78 0.00 2.02 -5.13 0.00 b11 0.00 --fixed-b12 0.324 0.0746 4.35 0.00 0.0665 4.87 0.00 b13 0.554 0.123 4.52 0.00 0.115 4.81 0.00 b21 0.00 --fixed-b22 -0.128 0.617 -0.21 0.84 * 0.580 -0.22 0.83 b23 -0.317 0.678 -0.47 0.64 * 0.666 -0.48 0.63 Vn1 = Vn2 = V3n = WS 13/14 0 5:57 + 0:324 d2;n 10:4 + 0:554 d3;n 0:128 sn 0:317 sn Einführung in Verkehr und Logistik Commute-to-school mode choice modelling for students located in Dresden, Saxony. 8 Choice set 8 1 > > < i= 2 > 3 > : 4 8 See foot bike transit car [MTH08]. * * 47 / 56 WS 13/14 Einführung in Verkehr und Logistik 48 / 56 Exogenous Variables I Alternative-specific constant (k = 0) I Distance between school and student’s location in [km] (k = 1) I Car availability (k = 2) zni2 = 1 0 Car always available else I Season=weather condition (k = 3) 1 Winter=bad weather zni3 = WS 13/14 0 else Einführung in Verkehr und Logistik Estimation results (N = 9300) Mode foot (i = 1) Variable (Coeff.) ASC (10 ) Distance (11 ) Car-avail (12 ) Season (13 ) bike ASC (20 ) (i = 2) Distance (21 ) Car-avail (22 ) Season (23 ) transit ASC (30 ) (i = 3) Distance (31 ) Car-avail (32 ) Season (33 ) Null log-likelihood: -12 892.50 Likelihood ratio test: 16200.80 Estimate SE 10.774 0.234 -4.376 0.183 -5.279 0.1634 -0.591 0.2385 6.570 0.190 -0.904 0.033 -4.772 0.184 -2.081 0.147 4.477 0.171 -0.052 0.017 -5.553 0.143 -0.489 0.135 Final log-likelihood: Rho-square: 0.628301 t-statistic 46.042 -37.008 -3.609 -22.129 34.596 -27.365 -22.599 -14.160 26.210 -2.966 -38.870 -3.621 -4792.14 49 / 56 WS 13/14 Einführung in Verkehr und Logistik 50 / 56 Utility Functions I Car-availability ”always” (zni2 = 1) I Summer term (zni3 = 0) WS 13/14 Einführung in Verkehr und Logistik I Car-availability ”not always” (zni2 = 0) I Summer term (zni3 = 0) 51 / 56 WS 13/14 Einführung in Verkehr und Logistik 52 / 56 How to compute utility values? I Distance to school 0.5 km (zni1 = 0:5) I Car is always available (zni2 = 1) I Summer term (zni3 = 0) Vn1 = 10; 774 Vn2 = 6; 570 Vn3 = 4; 477 Vn4 = 0 WS 13/14 4; 376 0; 5 0; 904 0; 5 0; 052 0; 5 5; 2796 1 = 3; 307 4; 772 1 = 1; 346 5; 553 1 = 1; 102 Einführung in Verkehr und Logistik Corresponding choice probabilities I Distance to school 0.5 km (zni1 = 0:5) I Car is always available (zni2 = 1) I Summer term (zni3 = 0) Pn1 = Pn2 = Pn3 = Pn4 = e 3;307 e 3;307 + e 1;346 + e e 1;346 e 3;307 + e 1;346 + e e 1;102 e 3;307 + e 1;346 + e e0 e 3;307 + e 1;346 + e 1;102 + e0 = 0; 841 = 84; 1% 1;102 + e0 = 0; 118 = 11; 8% 1;102 + e0 = 0; 010 = 1; 0% 1;102 + e0 = 0; 031 = 3; 1% 53 / 56 WS 13/14 Einführung in Verkehr und Logistik 54 / 56 Transit choice probabilities WS 13/14 Einführung in Verkehr und Logistik Literature I M. Ben-Akiva and S.R. Lerman. Discrete choice analysis, theory and applications to travel demand. MIT Press, Cambridge, MA, 1985. M. Bierlaire. An introduction to BIOGEME (Version 1.8), 2008. K. Haase, G. Desaulniers, and J. Desrosiers. Simultaneous vehicle and crew scheduling in urban mass transit systems. Transportation Science, 35(3):286, 2001. M. Klier and K. Haase. Line optimization in public transport systems. In J. Kalcsics and S. Nickel, editors, Operations Research Proceedings 2007, pages 473–478. Springer, 2008. S. Müller, S. Tscharaktschiew, and K. Haase. Travel-to-school mode choice modelling and patterns of school choice in urban areas. Journal of Transport Geography, 16(5):342–357, 2008. 55 / 56 WS 13/14 Einführung in Verkehr und Logistik Literature II A. Schöbel. Line planning in public transportation: models and methods. OR Spectrum, 34:491–510, 2012. E. Train, Kenneth. Discrete choice methods with simulation. Cambridge University Press, 2003. 56 / 56