Anda di halaman 1dari 6

Developing Fuzzy Route Choice Models Using Neural Nets

Yaser E. Hawas

Assistant Professor. Civil Eng. Department, C’niied Arab Emirates University


AI-Ain, UAE, P.O. Box 1755, Ph.: 971-37051699. E-mail:y.hawas@uaeu.ac.ae

Abstract Henn [16] developed a fuzzy route choice model to


accommodate the uncertainties of the drivers’
Thispaper discusses the calibraiion methodology of a behaviour. The model is compared with a stochastic
Neuro-Fuzv logic for route choice behaviour discrete choice logit model. The effect of the ATIS
modelling. The Neuro-Fuzzy refers io the recent information is modelled as a modification of the
trend of logics that couple the traditionaljiuzy logic probability (uncertainty) that the traveller perceives
structure with mural nets training capabilities for regarding the predicted route cost. Kuri and Pursula
knowledge bace and parameters seiiings. The J?AUY [17] compared between logit-type random utility
logic accounis for the various foctors of potential models and fuzzy logic.
effect on the route choice utility perceived by the
traveller. The struciure of the jiuzy logic, the This paper utilizes to the so-called “rule-based” fuzzy
calibration of the membership firnctionr, and the technologies. Nearly, all recent fuzzy applications in
composition of the knowledge base are discussed in the area of transportation engineering are based on
details. Logic training is based on data extractedfrom this methodology [15,16,17]. The key benefit of
a faciorial experimental design model. “rule-based” fuzzy logic is that it enables describing
the system with simple “if-then” relations. As such,
simpler solutions can be obtained with less design
1. Introduction time, and use of engineering knowledge can be easily
incorporated to optimise the system performance.
Route choice behaviour models represent the core of While this is certainly the advantage of fuzzy logic, it
the simulation models to support the Advanced is a major limitation. In many applications,
Traveller Information Systems (ATIS). Hawas [l] knowledge that describes a specific system status can
provides a detailed overview on the state-of-practice only be gathered by substantial exploration of huge
in the area of data collection, modelling, calibration data sets. The “if-then” rules (knowledge base) are
and validation of route choice models. commonly developed intuitively; through data
exploring and arbitrary simple reasoning arguements
Route choice modelling has been traditionally - needless to comment on the major effort involved.
modelled by using random utility modelling As such, very limited size knowledge bases were
approaches; mostly Logit or Probit models [2-71. developed and very few factors (primarily travel
Econometric methods, analysis of variance, cross time) were considered. Several factors of high
tabulations, regression techniques, and probabilistic potential influence on route choice were not
models were also widely used in studying and considered to simplify the knowledge bases.
modelling the route choice behaviour [8,9]. It can be Furthermore, to the author’s knowledge, no specific
argued that these models may not accurately capture methodology was adopted or suggested for the
the varying driver’s perception of the route trafic systematic calibration and verification of such fuzzy
states. Their distribution functions of the travellers’ logic knowledge bases.
behaviour (Gumbel’s or Normal distributions) may
not be valid, and as such these models’ applicability Recently, operational research scientific committees
might be limited. Laboratory simulation experiments were motivated to integrate the neural nets training
[2,6,10,11, 12,131, mail surveys [5,7], virtual reality capabilities into the development stages of the fuay
simulators [9], and internet-based graphical control logic. Neural nets would act as the agent for
simulators 1141 were utilized for data collection. deriving the if-then inference engine from the huge
data set. The integrated Neuro-Fuzzy logic have the
Fuzzy logic was recognized to be a very effective abilities to learn from data sets, derive its own
modelling approach for the route choice complex inference engine, and optimise its parameters.
process [15,16,17]. It can be used to overcome the
limitations of the distribution functions inhibited in
the conventional modelling approaches. Twdorovic 2. Fuzzy Route Choice System
et al [15] presented a fuzzy logic model for route
choice. The model’s knowledge base is developed The input variables are the factors of potential
using simple reasoning arguements. Data to test this influence on the driver’s perception of the route
algorithm were obtained from a computer simulation.

71
utility. In this paper, six factors were identified for 1. Prior Choice Route Utility Stage (PCS): this stage
each route alternative: calculates the route utility as perceived by the
traveller before the actual trip is made (when
1. Travel Time (numeric) helshe receives the information from any specific
2. Highway Classification (categorical) ATIS information source, S, at time, I). The input
3. Queuing Time (numeric) to this module is the matrix of the received factor
4. Familiarity (categorical) levels of all the mutes, Fj,s. The output of this
5. HighwayRavement Condition (categorical)
6. Speed (numeric) module is the vector of the prior routes’ utilities,
U;,s.The subscript P refers to the Prior utility.
Each numeric factor is regarded as a continuous real
number with some range (e.g. speed ranges from 0 to Initially, U;,, is set equivalent to the utilities
120 !un/hr). The range constitutes a set of smaller calculated at time f - I , (at f =0, set to zero)
intervals, each of which corresponds to a linguistic
term as shown in Table 1. The categorical factors are
discrefe integer numbers corresponding to linguistic 2. Following Choice Route Utility Stage (FCS): this
terms. At any time f, the ATIS information source, S, module calculates the route utility as actually
provides the traveller with a value (level) for each perceived by the traveller. The input to this
factor. Each level corresponds to one or more module is the matrix of acfually experienced
linguistic term (e.g. speed of 90 M r corresponds to factor levels, F;,,, The output of this module is
medium and high) with varying probabilities. the vector of actually perceived routes’ utilities,
Table 1. Term Definitions of the Route Choice
U>,s.The subscript A refers to the Actual utility.
Modelling Fuzzy System Initially,U>,sset is set equivalent to the utilities
Unit Min Max Linguistic
Variable Name Terms calculated at time f-1, U;; (at I =0, set to zero).
Travel Time Percent 100 650 Free
Light 3. Reliability Level (Experience) Stage (RLS): this
Medium module updates the routes’ reliability levels
High perceived by the travellers; to estimate the
travellers’ degree of confidence in the information
Highway Class 0 3 Freeway provided by S a t time 1. The route reliability level
AWP ’ is only updated if the prior perceived utility does
AWOP not match the acfually experienced one. The
input to this module is the vector of the difference
Queue Time minutes 0 25 Light between the prior and the actual utilities, dug.
Medium
Heavy The output is the vecfor of the route reliability
levels, R; . The dimension of this vector is equal
Familiarity 0 1 Familiar to the number of routes; each element represents
Non-Familiar the reliability level associated with a single mute.
hitially, R: set is set equivalent to &-I (at I
Speed kmihr 0 120 Low
Medium =O, set to unify vector).
High
4. Route Choice Decision Stage (RCS): This module
Highway/ 0 3 Poor emulates the route choice decision-making
Pavement Acceptable process. The input to this module includes
Condition Good primarily the most recent reliability vector, RA-’,
together with the mutes traffic states (factors’
Route Utility 0 450 VeryLow
Low levels), F;,,.The output of this module is the
Low-Medium vector of final route utility, U;,,. The subscript
Medium
Medium-High F refers to the Final utility.
Hieh
Very High This paper focuses on the development and
’’ART: with signal progression
Arterid
AWOP: Arterial without si& progression
calibration of the PCS and FCS modules. The logics
of two modules are identical (as shown in figure 1);
only the input values are different as previously
The logic system constitutes four various stages explained.
(modules). The role of each stage is explained below:

12
Factorial Design Model's Estimated Utility, U=
Mean Effect
C255.96
Factor Effects
.18.74A1+13.02 A2-2.01A3-4.71Ahc2.0OAs
+27.06Bi+24.15B2-0.84Bs
+14.4OC]+7.5OC2
+33.66Di+8.27D2
-14.06E
I C -12.94Fi+1.74F2
-19.65G1+2.39Gz
Interaction Effects
+O.~~AIB~+~.O~A~B~-~.~~A~BI+~.~~A&-O.~~ASBI
t0.76A1B2+2.48A2B2-l.95A3B2+1.63A&-0.85AsB2
-0.72AIB3-0.25A2B3-0.1 7A3B3+2.4OA&- 1.19AsB3
-1.01AICl+2.61AzC1+2.05A~C~-3.48hC~-0.1 lAsCi
+1. ~ ~ A I C ~ + O . ~ ~ A ~ C ~ - O . ~ ~ ~ A ~ C Z - ~ . ~ O & C ~ - O . ~ ~ A S
-~.~OAID~-~.~~AZDI+O.~~A~DI+O.~~A~~~+~.O~
-3.80A~D2-0.90A2D2+0.26A3D2+1.83~2+1 .75A5D2
r +OSOAlE+l. ~ ~ A ~ E - ~ . ~ ~ A ~ E - O . ~ ~ U + ~ . S O A S E
+2.97AiF1-1.24AzF1-2.86A3Fi+3.9O%F,+3.46A91
-2.18AIF2+1.22A2F2+1.62A~F2-0.089A&-2.65AsF2
Lwia*hrh -1.16AIGl+l.85A2G1+5.53A~G1-4.81&G1-2.16AsG1
+0.95A,G2-1 . ~ Z A ~ G ~ - O . ~ ~ A ~ G ~ - O . ~ ~ & G ~ + O . S ~ A S G ~
Nlmalow w
+4.17BiC1-1.34B2C1-1.66B3C1
-2.14B I C ~ + O . ~ ~ B Z C ~ + O . ~ S B ~ C ~
Fdlowlrg LMce Unlity -1.92B,E+0.42B2E+0.97B,E
+O.~OCIDI-O.O~SC;DI
+1.5OC1D2-1.42GD2
Ac!a~sl RoUteFaLtDIs, Pa -0.94CiE -0.2OC2E

- K&+WCadidM:Crwwkd
F.mli.yCIapaU
Achld
RDUh
FCS
~ t i i i t y ,U L
where:

A: City area-Age Factor


............
Fig. 1. PCS and FCS stages of the fuzzy logic for A,: NC-Y Non-congested area and young age
route choice behaviour system A2: NC-M Non-congested area and medium age
A,: NC-0: Non-congested area and old age
3. Data for Fuzzy Logic Training &:C-Y Congested area and young age
AS:C-M Congested area and medium age
Six participant groups were targeted for collection of Aa: C - 0 Congested area and old age
data needed for the calibaration. These groups were B: Travel Time
selecetd to partially capture the effect of socio- B,: Free B2: Light
demographic characteristics on the route choice B3: Medium B4: High
behavior. The participants are grouped based on the C: Highway Class
common M c characteristics (congested, not C,:Freeway
congested) in the city area they reside and their age. C2: Arterial with signal progression: AWP
Travellers residing congested areas demonstrate more C,: Arterial without signal progression: AWOP
tolerance with high congestion levels than those D: Queue Time
residing non-congested areas, and as such their D,: Light D2: Medium D,: Heavy
perceptions of the congestion levels would differ. E: Familiarity
Similarly, older age groups demonstrate higher E,: Familiar h:Non-Familiar
tolerance levels to congestion. Each participant was F: Highway/ Pavement Condition
asked to allocate a score (weight) for each F1: Poor F2:Acceptable F3: Good
factorfievel. These scores were then utilized to G: Speed
estimate the mute utility associated will all GI: Slow G2:Medium GI: High
factornevel combinations. The score-estimated
utilities corresponding to various factordlevels were The factorial experimental design model is used to
utilized to develop the factorial aperimenfa1 design generate the data required for the training and
model below. The details of data collection, and calibration of the fuzzy logic. Training data is
calibration of the factorial experimental design model generated in six different buckets representing the
can be found in [l]. various participant groups. Each data bucket
constitutes 848 data records. Table 2 shows a sample

73
of the data bucket of the C-0 participants’ group.
The various data buckets were then used to develop a
separate knowledge base for each patkipants’ group.

Table 2. Sample Training Data for Route Utility I


Fuzzy Logic ( G O Group)

Factor Values Estimated 1 ,Z


,‘.,.
1 ;
-I i“ i .; .i -i i - 1 - / L

Route --YID-

B C D E F G Utility*
130 0 3 0 0 30 356.1286
350 0 3 0 0 30 222.4297
130 1 3 0 0 30 324.6624
350 1 15 1 1 60 167.1657
130 2 15 1 1 60 193.9634
200 0 8 1 0 90 298.8737
230
350
0
0
8
8
1
1
0
0
90
90
275.8522
227.078
* V a l u s mimatedwing h e Factorial Expcn”tal Design Model
i
l
f

~
.

----
l , t , ,
*
. .

Fig. 2. Numerical Factors’ Levels Identification


t l
I
z #

w-.Mlb.(D*p1
I
*

Scale Sheet.
5. Neuro-Fuzzy Logic Development and
Training

The membership functions of the numerical factors


were estimated using group discussion surveys [l]. A
total of 60 students, faculty members and stafi
equally representing various residence locations and
ages participated in these group discussion surveys.
Each participant was asked to identify on some i Travel Time (ratio of travel tine to mtumumtrave tunc
numerical scales bow does he/& perceives the
linguistic terms describing the levels of the numerical
factors (namely, travel time, queue time and speed);
for example, free travel time: 10-12 minutes, light
travel time: 12-15 minutes, etc. Figure 2 illustrates a
sample of the surveying sheet for this purpose. These
data are then used to set the definition points of the
membership functions of the linguistic terms as
shown in figure 3.

Table 3 shows a sample of IF-THEN rules of the C-0


group knowledge base. Initially, the rules formed
exhaustively; all possible arrangements of
input/output ferns. Each rule is assigned a degree of
support, DOS,representing the weight of the rule. The
DOS ranges from m.10 to one; zero indicates a non-
valid rule, and one indicates a valid rule. Initially, the
rules formed exhaustively, and the DOS’S are set Fig. 3. Membership functions of the numerical
randomly in the range [OJ]. The Neuro-Fuzzy factors
learning capability is used to calculate the DOS’S.
Table 3. Sample of the Rules of the Route Utility
The rules with DOS close to zero are considered non-
valid rules and could be eliminated from the IF-THEN Block
IF THEN
knowledge base. The training data from the factorial
experimental design model are utilized for developing Route
E C F D G B DOS Utility
the knowledge base by calibrating the DOS of the
Familiar Freeway Pwr Light Low Medium 0.59 Medium
rules, as well as calibrating the membership functions
of the route utility terms. Familiar F m y Poor Light Low High 0.80 Medim
Familiar Freeway Poor Light High F m 1.00 High

74
The Fuzzy-Tech software [18] is utilized in the Acknowledgement
development of the PCS and FCS modules. Data
training is carried out using a neural net iterative This research has been funded by the Research
algorithm based on the technology of Fuzzy Council at the UAE University under grant numbers
Associative Maps or Memoly @AM) [19]. The 04-7-11/2000 and 05-7-11/2001.
DoS’s are considered the training parameters. The
idea is to find the DoS’s that best match the training
data and the fuzzy logic derived utilities.

The algorithm calculates an error gradient by slightly


changing the DOS. At any iteration, the fuzzy logic
utilizes the input data of the training records to
calculate the route utility. Then, an error vector of
the difference between the calculated utilities and the
training data utilities is estimated. A heuristic
procedure is used to identify the rule most suited for
influencing such error vector. The DOSof this rule is
modified before another iteration is carried out. Upon
fulfilling any of the pre-specified stopping criteria the
training process is terminated.

Initially, the definition points of the route utility terms


are set arbitrary as shown in figure 4.a. Each term is
defined by a single base value that corresponds to
membership value of 1 (e.g. low route utility base is
set to 112 in Figure 4.a). The Neuro-Fuzzy uses the
term base for adaptation; a slight change of the base
will be accompanied by a horizontal shift of the term
maximum as well as a shift for the neighbouring
terms defmition points as indicated by the dashed set
shown in Figure 4.b. The base shifting is carried out
for a single term at a time, and simultaneously with
the DOS learning so as to achieve minimum error tm m 101 a
l
vector between mining data and fuzzy logic derived (1) Row Mlm
utilities. Figure 4.c shows the utility membership Fig. 4. Route utility defuzzification learning
functions upon terminating the learning procedure.

5. Route Utility Model Validation


Model validation is conducted through experimental
templates that illustrate for each traveller three routes
with various factor levels. The numeric factor levels
of the travel time, queuing time and speed were given
(not linguistic terms). 500 templates were prepared to
cover the various factor levels. Figure 5 shows a
sample template for validating the survey data and the
modelling approach. Each participant was randomly
assigned 20 templates and asked to select one of the
three routes shown in the template. Figure 6 shows
the contour maps of percentage agreement between
model and ochral route choice decisions; the darker
the shaded area, the higher the discrepancy between
the model and the actual decisions. The percentage
agreement ranges fiom 84 to 94%. These results
indicate that the developed fuzzy logic is quite
effective in capturing the route choice behaviour.

Fig. 5. Sample template for validating the survey


data and the modelling approach.
Studying the Effect of Advanced Traffic
Information on Drivers’ Route Choice”.
Transportation Research, Part C, Vol. 5 , No 1,
pp 39-50, 1997.
Vaughen, K.M.; Kitamura, R., and Jovanis,
P.P. Experimental Analysis and Modeling of
Advice Compliance: Results h m Advanced
Traveler Information System Simulation
Experiments. Transportation Research
Record, 1485,pp 18-26, 1995.
Katsikopoulos, K.V., Duse-Anthony, Y.,
Fisher, DL., Duffy, S.A. “The Framing of
Drivers’ Route Choices When Travel Time
Information is Provided Under Varying
Degrees of Cognitive Load”. Human Factors,
voi. 42, NO. 3,pp 470-481,2000.
.
K” *i* e-3 ”.c CI UI
[lo] Adler, JL. “A Conflict model and Interactive
Simulator (FASTCARS) for Predicting En-
Otl*GrmP
Route Driver Behavior id Response to Real-
Fig. 6. Percentage agreement between the model- Time Traffic Condition Information”.
estimated and actual route choice decisions. Transportation, Vol 20, No. 2, pp 83-106,
1993.
[ l l ] Chen, P.S. and Mahmassani, H.S. “Dynamic
References Interactive Simulator for Studying Commuter
BehaviorUnder Real-Time Traffic Information
Hawas, Y.E. “Development and Calibration of Supply Strategies”. Transportation Research
Route Choice Behavior Models: Factorial Record, 1413, pp 12-21,1993.
Experimental Design Approach”. Paper [I21 Koutsopoulos, H.N., Lotan, T., and Yang, Q.
Submitted for Publication in the “A Driving Simulator and Its Application for
Transportation Research, Part C, 2001. Modeling Route Choice in the Presence of
Vaughn, K.M., Ahdel-Aly, M.A., Kitamura, Information”. Transportation Research, Part
R., Jovanis, P.P., Yang, H., Kroll, N.E.A., C, Vol2, No. 2, pp 91-107, 1994.
Post, R.B., and Oppy, B. “Experimental [I31 Bonsall, P., Firmin, P., Anderson, M., Palmer,
Analysis and Modeling of Sequential Route I. And Balmforth, P. “Validating the Results
Choice Under an advanced Traveler of Route Choice Simulator”. Transportation
Information System in a Simplistic Traffic Research, Part C, Vol. 5 , No. 6, pp 371-387,
Network“. Transportation Research Record, 1997.
1408, pp 7542,1993. [I41 Ozhay, K., Datta, A., and Kacbroo, P.
Srinvasan, K.K., and Mahmassani, H.S. “Modeling Route Choice Behavior Usin
“Modeling Inertia and Compliance Stochastic Learning Automata”. Proc. 8b
Mechanisms in Route Choice Behavior Under Annual Meeting of the Transportation
Real-time Information”. Transportation Research Board, Washington, D.C, 2001.
Research Record. 1725, pp 45-53,2000, 1151 Teodorovic, D., Vukanovic, S., and Obradovic,
Abdel-Aly, M.A., Vaughen, K.M., Kitamura, K. “Modeling Route Choice with Advanced
R., Jovanis, P.P., and Mannering, F.L. Traveler Information by fuzzy Logic”.
“Models of Commuters’ Information Use and Transportation PIanning and Technologv, Vol.
Route Choice: Initial Results Based on 22, No. 1, pp 1-25, 1998.
Southern California Commuter Route Choice [I61 Henn, V. ‘‘Fuzzy Route Choice Model for
Survey”. Transportation Research Record, Traffic Assignment”. Fuzzy Sets and System,
1 4 5 3 ; ~46-55,1994.
~ Vol. 116, No. 1, pp 77-101,2000.
Emmerink, R.H.M., Nijkamp, P., Rietveld, P., [I71 Kuri, I. and Pursula, M. “Modeling Car Route
Van Ommeren, J.N. “Variable Message Signs Choice with Non-Dynamic Stated Preference
and Radio Traffic Information: An Integrated Data: A Comparison Between Fuzzy Similarity
Empirical Analysis of Drivers’ Route Choice and Random Utility Models”. Proc 8@
Behavior”. Transportation Research, Part A. Annual Meeting of the Transportation
Vol. 30,No. 2, pp 135-153, 1996. Research Board, Washington, D.C, 2001.
Lotan, T. “Effects of Familiarity on Route [IS] INFORM. FuuyTECH 5.52: User’s Manual.
Choice Behavior in the Presence of GmbH I Inform Software Corporation,
Information”. Transportation Research, Part Germany, 2002.
C, Vol. 5, No. 3, pp 225-243, 1997. [I91 Ross, T. I. Fuzzy Logic with Engineering
Abdel-Aty, M.A., Kitamura, R. and Jovanis, Applications, McGraw-Hill, Inc. 1995.
P.P. “Using Stated Preference Data for

76

Anda mungkin juga menyukai