Академический Документы
Профессиональный Документы
Культура Документы
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 1, FEBRUARY 2014
I. I NTRODUCTION
Manuscript received October 14, 2012; revised March 10, 2013 and June
4, 2013; accepted July 16, 2013. Date of publication August 21, 2013; date of
current version January 31, 2014. This work was supported by the National Key
Basic Research Development Program of China under Grant 2012CB725405.
The Associate Editor for this paper was L. Li.
P. J. Jin is with the Center for Transportation Research, Department of
Civil, Architectural, and Environmental Engineering, The University of Texas
at Austin, Austin, TX 78701 USA (e-mail: jjin@austin.utexas.edu).
D. Yang is with the School of Transportation and Logistics, Southwest
Jiaotong University, Chengdu 610031, China (e-mail: yangd8@gmail.com).
B. Ran is with the School of Transportation, Southeast University, Nanjing
210096, China, and also with the Department of Civil and Environmental
Engineering, University of Wisconsin, Madison, WI 53706 USA (e-mail:
bran@wisc.edu).
Digital Object Identifier 10.1109/TITS.2013.2273872
1524-9050 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
JIN et al.: REDUCING THE ERROR ACCUMULATION IN CAR-FOLLOWING MODELS WITH TRAJECTORY DATA
1 yi yi
MAPE =
yi +
N
i
1 yi yi 2
RMSPE =
N i
yi +
1 (
yi yi )2
N
i
U =
1
1 (y 2 )
2) +
(
y
i
i
N
N
i
i
|
yi yi |
J= i
|yi |
(4)
(5)
(6)
min
f (, Y, Y ) + C(Y, Y )
sn (t)
s.t.
ri () 0,
sj (Y , ) 0,
ln
i = 1, . . . , m
j = 1, . . . , n
(3)
The calibration of a car-following model is a typical nonlinear optimization problem. The objective function of the problem is an MOE (measure of effectiveness) function reflecting
the difference between model outputs and the ground truth data.
The constraints include the valid ranges of model parameters
and physical constraints on the vehicle dynamics (e.g., maximal
velocity, acceleration, and minimal spacing). The calibration
problem based on vehicle trajectory data can be generalized as
the following:
149
(1)
xn1 (t)
xn+1 (t)
vn (t)
an (t)
(7)
(8)
(9)
(10)
150
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 1, FEBRUARY 2014
(11)
(13)
(14)
fxn
0
f vn
an (t)
an (t + )
vn (t + ) = 0 1 + fvn
fxn vn (t)
2
2
xn (t + )
xn (t)
0 + 2 fvn 1 + 2 fxn
an (t)
:= A vn (t) .
(20)
xn (t)
Then,the three eigenvalues of A are 0, and (1/2)fvn
( /2) (fvn + (fxn /2))2 + 4fxn + (fxn 2 /4) + 1. Therefore, for errors to decay over time, the following condition
should hold:
2
2
1
fxn
fxn
< 1.
fv
+
1
+
+
4f
+
f
vn
xn
2 n
2
2
4
(21)
where fxn , fvn , and fan are partial derivatives with respect
to the location, velocity, and acceleration. Furthermore, to
focus on the properties of the error propagation, we assume
no systematic errors in car-following models, that is, an (t +
) = f (xn (t), vn (t), an (t)). Then, (14) can be simplified as the
following:
Substituting fvn and fxn with (22), the error stability condition
becomes the following:
2
2
(23)
+ 1 < 1.
4
2
2
4
2
< 1.
(24)
1+
2
vn (t + ) = vn (t) + an (t + )
(17)
1
xn (t + ) = xn (t) + vn (t) + 2 an (t + ). (18)
2
Combining (15), (17), and (18), the following error dynamic
system can be obtained:
fvn (t)
fxn (t)
fan (t)
an (t + )
vn (t + ) = fan (t) 1+ fvn (t)
fxn (t)
2
2
2
xn (t + )
+ 2 fvn (t) 1 + 2 fxn (t)
2 fan (t)
an (t)
an (t)
vn (t) := A vn (t) . (19)
xn (t)
xn (t)
fvn = 0
and
fxn = .
(22)
(25)
JIN et al.: REDUCING THE ERROR ACCUMULATION IN CAR-FOLLOWING MODELS WITH TRAJECTORY DATA
151
fxn = OV (xn ).
(26)
OV
(x
)
+
4OV (xn )
n
2
2
2
2
OV (xn ) + 1 < 1. (27)
4
Whether or not the two nonzero eigenvalues have an imaginary
part or not, (27) will be different inequalities from the modelstability condition [2], i.e.,
OV (xn ) + /2 > 0.
(28)
(29)
N
1
I (i)
N i=1 an (t)>Ta
(30)
152
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 1, FEBRUARY 2014
EVR = CEVR
CR = CCR
N
1
I (i)
(i)
N i=1 vn (t)<0 or vn (t)>Tv
(31)
N
1
I (i)
(i)
N i=1 xn (t)>xn1 (t)
(32)
(33)
N
1
(wi |
yi yi |)
N i=1
1
(wi (
yi yi ))2
N
wUi =
1
N
((wi yi )2 ) +
1
N
(34)
(35)
((wi yi )2 )
(36)
(37)
JIN et al.: REDUCING THE ERROR ACCUMULATION IN CAR-FOLLOWING MODELS WITH TRAJECTORY DATA
153
TABLE I
C ALIBRATION R ESULTS
Fig. 2.
where the weights are determined by visual inspection of accumulative plots (see Fig. 2 for an example). It should be noted
that the combined MAE only serves as the initial screening index to select candidate objective functions for further evaluation
through the accumulative error analysis.
To evaluate the error accumulation, the simulated data are
further divided into groups based on the number of iterative calculations made. Within each group, the similar adjusted MAEs
defined in (36) are calculated. The error accumulation can then
be analyzed by plotting the curves of the adjusted MAEs of
each group versus the corresponding number of iterations for
the acceleration, velocity, and location.
VII. R ESULT A NALYSIS
A. Combined MAE Evaluation for Training Results
The shaded cells in Table I highlight the smallest five adjusted MAEs obtained for each algorithm. The bold results are
the best adjusted MAEs obtained for each model. The performance of Gipps and FRESIM are better than the other three
models with lower invalid rates and smaller combined adjusted
MAEs for the US101 data set. According to the adjusted MAEs,
WxMAE has the best overall error performance for GM and
Bandos algorithm. The xCR and sCR achieve the optimal
overall error performance for Gipps model. The traditional
velocity based MOEs, the vMAE and vTheil have the best
overall performance for FRESIM and IDM models. However, it
may be caused by the error characteristics of FRESIM and IDM
models. FRESIM strictly enforces the zero-crash rate, which
achieves the best error accumulation performance regardless
of which objective function used, whereas IDM has the worst
invalid rate results among all five algorithms. Furthermore, in
general, WxMAE has the best overall performance as its error
performance falls into the top five for all five algorithms. Based
on the overall performance, eight MAE functions, namely,
vMAE, sMAE, vTheil, xTheil, vEVR, xCR, WxMAE, and
WsU, are chosen for further evaluation of their error accumulation performance.
All the calibrated model parameters and their valid ranges
can be found in Table II. Note that the model parameter ranges
are selected based on the previous literature and range for
the parameters for the penalty terms are determined based on
their acceptable ranges, in which the error statistics and the
penalty terms are still comparable. The standard deviations of
the resulting optimal parameter values from the multiple runs
of the GA are also reported. Most of the model parameters are
concentrated around the optimal values except for the sensitivity variable in the GM model and the two constant coefficients
154
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 1, FEBRUARY 2014
TABLE II
VALID R ANGES FOR PARAMETERS AND THE C ALIBRATED VALUES
Fig. 3.
Subplots in each row are the accumulative error plots for the
acceleration, velocity, and location, respectively. Fig. 2 shows
the error accumulation curves of the best four objective functions using traditional error measures. To further demonstrate
the impact of the error accumulation, we use the calibrated
model to conduct one-step car-following simulation, that is,
the results are obtained by executing each car-following model
once and the input data directly come from ground truth data.
Each point in Fig. 3 is the average one-step MAE for each
corresponding point in Fig. 2. As illustrated in Fig. 3, the
calibrated models are quite accurate in one-step simulation,
particularly in matching the ground truth velocities and locations with average errors about 1/3 and 1/20 of the multistep
simulation errors, respectively. Such significant differences in
simulation error demonstrate the significance of addressing the
error accumulation for multistep car-following simulation with
models calibrated from field trajectory data.
According to Figs. 2 and 4, among all five algorithms,
the Gipps and FRESIM model have the best accumulative
error performance regardless of the objective functions used.
Moreover, such phenomenon may be caused by their crash
avoidance mechanism and the restriction of the high-order
terms with the parameter b and Vlim . If we consider half the
values of Ta , Tv , and Tx are acceptable in simulation, Gipps
and FRESIM models can both hold for much more than 25
time steps, whereas the other three algorithms can only stay
within the error range in less than 510 steps. Both GM and
Bando models have clear trends of error accumulation among
all objective functions tested. WxMAE has less acceleration
and velocity error accumulation performance than the others.
For the Gipps model, the same, xCR, and WxMAE all returned
optimal location error accumulation results. The IDM model
has a trend different from the others in the accumulative error
plot. Although it suffers from fluctuating error accumulation
JIN et al.: REDUCING THE ERROR ACCUMULATION IN CAR-FOLLOWING MODELS WITH TRAJECTORY DATA
Fig. 4.
155
(38)
(39)
(40)
156
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 1, FEBRUARY 2014
vn (t)
vn (t)
bmax
vn (t + ) = min vn (t) + 2. 5amax 1
0.025 +
V
V
+
2
b2max 2 bmax 2 [xn1 (t) ln1 d xn (t)] vn (t) vn1
(t)/b0
(42)
s (vn (t), vn (t)vn1 (t))
vn (t)
1
(43)
v0
xn1 (t)xn (t)ln1 (t)
+ T vn (t) +
.
= s0 + s1
v0
2 a 0 b0
(44)
There are six coefficients, namely, a0 , b0 , T , s0 , s1 , and in this
model. However, according to the recommended settings in [9],
s1 = 0 and = 4. Therefore, only the former four parameters
need to be calibrated.
A PPENDIX II
C AR -F OLLOWING P LATOON E XTRACTION A LGORITHM
FOR F ULL T RAJECTORY DATA
The platoon extraction algorithm take NGSIM vehicle trajectory data and generate the car-following platoons according
to predefined platoon size Mp , in this paper Mp = 2. The
algorithms identify four major microscopic events on each
lane, including vehicle sinks, vehicle sources, vehicle disengagements, and vehicle engagement. Vehicle sinks and sources
refer to vehicle leaving and entering the current lane due to
(41)
JIN et al.: REDUCING THE ERROR ACCUMULATION IN CAR-FOLLOWING MODELS WITH TRAJECTORY DATA
157
Da Yang received the B.S. and M.S. degrees in logistics engineering from Southwest Jiaotong University,
Chengdu, China, in 2007 and 2009, respectively,
where he is currently working toward the Ph.D.
degree in transportation planning and management.
From August 2010 to August 2012, he was a
Visiting Scholar with the University of Wisconsin,
Madison, WI, USA. His research interests include
car-following modeling and heterogeneity in carfollowing behavior.