Journal of Transportation Technologies, 2013, 3, 260-265

http://dx.doi.org/10.4236/jtts.2013.34027 Published Online October 2013 (http://www.scirp.org/journal/jtts)

Optimal Path Finding Method Study Based on Stochastic

Travel Time
Zhanquan Sun, Weidong Gu, Yanling Zhao, Chunmei Wang
Key Laboratory for Computer Network of Shandong Province, Shandong Computer Science Center, Jinan, China
Email: [email protected]

Received July 22, 2013; revised August 25, 2013; accepted September 18, 2013

Copyright 2013 Zhanquan Sun et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

ABSTRACT
Finding optimal path in a given network is an important content of intelligent transportation information service. Static
shortest path has been studied widely and many efficient searching methods have been developed, for example
Dijkstras algorithm, Floyd-Warshall, Bellman-Ford, A* et al. However, practical travel time is not a constant value but
a stochastic value. How to take full use of the stochastic character to find the shortest path is a significant problem. In
this paper, GPS floating car is used to detect road sections travel time. The probability distribution of travel time is
estimated according to Bayes estimation method. The combined probability distribution of a feasible route is calculated
according to probability operation. The objective function is to find the route that has the biggest probability to arrive
for desired time thresholds. Improved Genetic Algorithm is used to calculate the optimal path. The efficiency of the
proposed method is illustrated with a practical example.

Keywords: Optimum Path; Stochastic Travel Time; Genetic Algorithm; Floating Car

1. Introduction rithm. Classical shortest path searching methods are

based on fixed path costs. However, practical travel time
Finding optimal paths in a given network is one of the
of a road segment is a stochastic variable. Shortest path is
most fundamental problems in network studies and has
a random variable also. Shortest path based on random
broad applications in the physical and social sciences and travel time has been studied [8-10]. The studied methods
in engineering, particularly in the fields of computer sci- are difficult to be used in practical because that the algo-
ence, operations research, and transportation engineering rithms are difficult and the probability distribution of
[1]. Static shortest path has been studied widely and road section travel time is difficult to be obtained.
many efficient searching methods have been developed, Travel time collection is the precondition of optimal
for example Dijkstras algorithm, Floyd-Warshall, Bell- path finding. Many kinds of collectors can be used to
man-Ford, A* et al. Dijkstras algorithm is a graph search collect traffic flow data, such as loop induce collector,
algorithm that solves the single-source shortest path microwave collector, video collector and so on [11]. The
problem for a graph with non-negative edge path costs, shortcoming of fixed collector is that the coverage is
producing a shortest path tree [2,3]. This algorithm is small and cost is expensive. It is difficult to obtain whole
often used in routing as a subroutine in other graph algo- roadway traffic state. Instantaneous traffic information
rithms. The Floyd-Warshall algorithm compares all pos- can be collected by the Floating Car Data (FCD) method.
sible paths through the graph between each pair of verti- It has become the efficient means to obtain real-time tra-
ces [4]. The Bellman-Ford algorithm is an algorithm that ffic information [12]. GPS floating car can collect travel
computes shortest paths from a single source vertex to all speed parameter. Different driver has different drive style.
of the other vertices in a weighted digraph [5]. Bellman- The driving speed corresponding to different driver may
Ford is based on the principle of relaxation, in which an be different under the same traffic condition. The travel
approximation to the correct distance is gradually re- speed is a stochastic variable. Average travel speed is
placed by more accurate values until eventually reaching often used to analyze traffic state. The stochastic charac-
the optimum solution. A* uses a best-first search and ter of traffic flow parameter is ignored. In this paper, the
finds a least-cost path from a given initial node to one stochastic character is considered in the optimal path
goal node [6,7]. It is an extension of Dijkstras algo- finding study. The probability distribution of travel time

Copyright 2013 SciRes. JTTs

 Z. Q. SUN ET AL. 261

has been widely studied [13]. In this paper, norm distri- d

bution is considered. t (3)
v
After the travel time probability distribution being de-
termined, how to find the optimal path based on stochas- 2.2. Travel Time Distribution Estimation
tic distribution is a difficult problem. In the optimal path
searching study, the least expected travel time is often A road section may include several floating cars in a
taken as the objection [14]. In the optimum problem, it sample time interval. The travel time of the road section
requires evaluating multiple integrals. Classic shortest based on each floating car can be calculated according to
path finding methods, such as Dijkstras algorithm, above introduced method. The probability distribution
Floyd-Warshall, Bellman-Ford, A* and so on, are not style is determined with hypothesis test method. The dis-
suitable to be used here. Some heuristic methods have tribution parameters of travel time are determined with
been adopted to resolve the optimal problem [15-17]. Bayes method. The estimation method is introduced as
Based on previous research, GA-based routing algorithm follows.
has been found to be more scalable and insensitive to Suppose a continuous random variable be X ,
variations in network topologies. It is taken as the most X ~ f . , , with continuous parameter space .
efficient one. However, many paths generated through The parameter s priori density is . . M X is the
cross and mutation are unreasonable. For resolving the observation space of X . Based on data points
problem, some improved genetic algorithms were pro- x1 , x2 , , xn , the Bayes theorem can be expressed as
posed. In this paper, the improved genetic algorithms follows.
proposed in [17] will be adopted to resolve the optimal l ; x1 , x2 , , xn
path problem. At last, an example is analyzed to illustrate x1 , x2 , , xn
the efficiency of the proposed method. l ; x1 , x2 ,, xn d
(4)
2. Travel Time Distribution Estimation where l ; x1 , x2 ,, xn is the likelihood function of
Based on GPS Floating Car the observations.
2.1. Travel Time Calculation Based on GPS When travel time of a road section obeys normal dis-
Floating Car tribution, the probability distribution can be denoted by

GPS floating car can provide localization data, speed, 1 x 2

f x exp (5)
travel direction and time information. Data collection 2 2 2

interval, which can be set according to practical require-
ment, is often set as 30 seconds. These data are the es- Given n data points D0 x1 , x2 , , xn , who obey a
sential source for intelligent transportation systems. In a normal distribution, the two parameters joint priori-dis-
certain time interval, a road section maybe includes sev- tribution according to Jeffreys method is
eral sample points of a floating car. Travel speed of a 1
road section can be calculated according to simple float- , (6)
2
ing car. The calculation method is described as follows.
Let the length of a road section be denoted d , the ve- The expected values of the updated parameters and
hicle speed at the ith sample point be denoted by vi and are
the number of sample point be denoted by p . Let t0 x x xn
D d 1 2

(7)
denote the first sampling point, and t1 denote last sam- n
pling point. Travel speed of the road section is calculated
0 D d

as follows.
1 2
t1
1
d t vt dt 1 p 1 v v 0

exp 2 i 1 xi d d (8)
n

v 0
r 0 tr 1 tr r 1 r (1) n 2
2
td td td 2
1 1
0 n 2 exp 2 2 i 1 xi d d
n 2
When the sample time interval is fixed, the formula
can be simplified as
According to above Bayesian method, the distribution
v0 v p of travel time can be determined.
r 1 vr
p 1

v 2 (2)
p 3. Distribution of Shortest Path
Travel time of a road section can be calculated, i.e. Let G S , E , T be a road network, where S is the

Copyright 2013 SciRes. JTTs

262 Z. Q. SUN ET AL.

vertices set, E is the link set, and T is the weight cult to obtain the analytic solution. Genetic algorithm is a
value set of E . Let R S1 S 2 S k denote a path from problem solving method based on the concept of natural
start node S1 to end node S n . The travel time of sec- selection and genetics. It has been quite successfully ap-
tion i, i 1, 2,, k is denoted by X i that obey normal plied in machine learning and optimization problems. It
distribution N i , i . The shortest path is denoted by is adopted to resolve the optimization problem. Classic
variable Y . It can be calculated according to the fol- genetic algorithm will generate unreasonable path. Im-
lowing equation. proved genetic algorithm is used to analyze the optimum
y x1 x2 xn (9) problem. The algorithm is introduced as follows.
The GA design involves several key components: ge-
The travel time of each road section is supposed inde- netic representation, population initialization, fitness
pendent. According to probability theory, the sum of function, selection scheme, crossover, and mutation. A
independent normally distributed variables is also nor- sample road network is shown as in Figure 1. The objec-
mally distributed. The distribution parameters are as fol- tion is to find the optimum route from source node 1 to
lows. destination node 12. The GA method can be described as
E Y E X 1 X 2 X n follows.
1) Genetic representation
E X1 E X 2 E X n (10) A routing path from source node to the destination
1 2 n node is encoded by a string of positive integers that rep-
resent the IDs of nodes through which the path passes.
D Y D X 1 X 2 X n Each position of the string represents an order of the
D X1 D X 2 D X n (11) node in the road network. A feasible route is from the
source node to the destination node without circle. Fig-

2 2 2
1 2 n
ure 2 is a sample from source node to the destination
The distribution of shortest path y can be described as node.
2) Population initialization
y 2 The initial population P is composed of a certain
1
f y exp
y
(12) number of chromosomes. The source node is taken as the
2 y 2 2
y
starting point of a chromosome p , which is constant. A
where feasible route is generated though selecting one of the
y 1 2 n neighbors provided that it has not been picked before. It
keeps doing this operation until it reaches to destination
and node. The population scale is determined according to
y2 12 22 n2 . practical requirement.
After the probability distribution being determined, the 3) Fitness function
probability of arriving at objection in time t is as fol- For a given solution, evaluate its quality, which is de-
lows. termined by the fitness function. In this paper, the fitness
function is the probability of arriving at destination in
pt p y dy
t
(13) time t . The probability distribution of a given solution

4. Shortest Path Calculation with

Genetic Algorithm
The optimum route is obtained through maximization
p t . It can be summarized as following optimization
problem.
y 2
1
max p t exp dy
t y

2 y 2 y2 (14)

sub to S1 S2 S k R
Figure 1. Road network.
where R is the feasible path set from source node to
destination node. 1 2 3 7 11 12
From the above optimization problem, we can see that
the object function is an integration function. It is diffi- Figure 2. A feasible path from source to destination node.

Copyright 2013 SciRes. JTTs

 Z. Q. SUN ET AL. 263

is calculated according to (13). method is used to forecast missing or deleted running

4) Crossover and mutation GPS floating car data.
The single-point crossover is supposed to exchange
partial chromosomes. With the crossover probability, 5.2. Travel Time Calculation
each time we select two chromosomes ri and rj for
Travel time is estimated with (3) according to each float-
crossover. They should possess at least one common
ing car. The probability distributions of travel time of the
node. The crossover process is shown as in Figure 3.
road section corresponding to each time section are esti-
After selection and crossover the population will undergo
mated with Bayesian method. The obtained distribution
the mutation operation.
parameters of each time section are listed in Table 1.
With the mutation probability, each time we select one
chromosome ri on which one gene is randomly selected
5.3. Optimum Route Calculation
as the mutation point. The mutation will replace the sub
path by a new random sub path. Mutation process is Use the improved GA method of section 4 to find the
shown as in Figure 4. optimum path. The parameters of GA are set as follows.
5) Selection The initial population number is set 5 and generated
After crossover and mutation, initial and new gener- chromosomes are shown in Figure 5. The crossover rate
ated chromosomes are used for selection. The chromo- and mutation rate are set 0.4 and 0.1 respectively. The
some with high fitness value has great chance to be cop- iteration threshold value is set 20. When t 10 , the op-
ied to the next generation. timum route is Ropt 1 2 3 7 11 12 whose
Repeat the process until terminating condition is probability is 0.97. When t 7 , the optimum route is
reached. Ropt 1 2 3 4 8 12 whose probability is
0.31.
5. Examples
5.4. Results Analysis
One year GPS floating car data of 3500 taxis are pro-
vided by Jinan Urban Public Passenger Transport Man- From the examples analysis results we can find that the
agement Services Center. In the aggregation of GPS optimum route based on stochastic travel time can be
floating car data, the upload interval is set 20 seconds. obtained with the proposed method. When the desired
The uploaded record items include vehicles running speed, travel time is different, the optimum route is different
direction, longitude, latitude et al. parameters. The lon- also.
gitude, latitude, direction parameters are used to locate
the vehicle position with map matching method. The pro- 6. Conclusion
bability of the traffic state is analyzed with the pro- posed
Optimum path based on random travel time is an impor-
method. The selected road network is shown as in Figure
tant problem. The optimum path finding method pro-
1. The time interval between two analyses is 5 minutes.
posed in this paper is based on stochastic travel time,
The traffic flow data from 8:00 to 8:05 are selected.
which can describe the real traffic situation properly. The
probability of arriving at destination point in desired time
5.1. Data Processing
is taken as the objection function. Classic optimum path
Firstly, we will make the map matching according to the finding methods, Dijkstras algorithm and A*, are not
GPS floating car data. The data that are invalid to ana- suitable to be used to solve the optimum path problems.
lyze traffic flow are deleted. They are recovered with In this paper, an improved genetic algorithm is adopted
forecasting method. In this example, exponent smoothing to solve the optimum problem. From the practical exam-
ple analysis results we can find that the proposed method
R1 1 2 3 7 6 10 11 12 is efficient in dealing with optimal path problem. GPS
R2 1 2 3 4 8 7 11 12
floating car has been widely adopted. It is possible to

R1 1 2 3 7 11 12 R1 1 2 6 10 11 12
R2 1 2 3 7 6 10 11 12 R1 1 5 6 7 8 12

Figure 3. Crossover of chromosome. R1 1 2 6 5 9 10 11 12

R1 1 2 3 7 6 10 11 12
R1 1 2 3 7 11 12
R1 1 2 6 10 11 12 R1 1 2 3 4 8 7 11 12

Figure 4. Mutation of chromosome. Figure 5. Initial population.

Copyright 2013 SciRes. JTTs

264 Z. Q. SUN ET AL.

Table 1. Probability distribution parameters of road network , .

2

1 2 3 4 5 6 7 8 9 10 11 12

1 0 1.5, 0.23 1.6, 0.16

2 1.8, 0.19 0 1.7, 0.28 100 1.4, 0.12
3 1.6, 0.24 0 2.1, 0.22 1.4, 0.12
4 2.2, 0.18 0 1.6, 0.20
5 1.5, 0.13 1.8, 0.31 2.2, 0.14
6 1.4, 0.15 1.9, 0.21 1.6, 0.21 2.7, 0.21
7 2.4, 0.33 1.7, 0.22 2.4, 0.33 1.6, 0.13
8 1.7, 0.26 2.5, 0.32 1.2, 0.21
9 2.1, 0.21 1.8, 0.31
10 2.5, 0.33 1.9, 0.16 1.8, 0.31
11 1.7, 0.14 1.8, 0.28 1.8, 0.31
12 1.3, 0.22 1.7, 0.30

Note: denotes that the two node are not connected directly.

apply the proposed method to practical system. With the for the Heuristic Determination of Minimum Cost Paths,
development of floating car, the data scale will become IEEE Transactions on Systems Science and Cybernetic,
Vol. 4, No. 2, 1968, pp. 100-107.
larger and larger. How to extend the proposed optimal
http://dx.doi.org/10.1109/TSSC.1968.300136
path finding method to large scale data is our future re-
[8] I. Chabini, Discrete Dynamic Shortest Path Problems in
search work.
Transportation Applications: Complexity and Algorithms
with Optimal Run Time, Transportation Research Re-
7. Acknowledgements cord, Vol. 1645, No. 1, 1998, pp. 170-175.
http://dx.doi.org/10.3141/1645-21
This work is partially supported by national youth sci-
ence foundation (No. 61004115), national science foun- [9] Y. Fan, R. E. Kalaba and J. E. Moore, Shortest Paths in
Stochastic Networks with Correlated Link Costs, Com-
dation (No. 61272433), and Provincial Fund for Nature puters and Mathematics with Applications, Vol. 49, No.
project (No. ZR2010FQ018) 9-10, 2005, pp. 1549-1564.
http://dx.doi.org/10.1016/j.camwa.2004.07.028
REFERENCES [10] E. D. Miller-Hooks and H. S. Mahmassani, Least Ex-
pected Time Paths in Stochastic, Time-Varying Trans-
[1] Y. Nie and Y. Y. Fan, Arriving-On-Time Problem: Dis- portation Networks, Transportation Science, Vol. 34, No.
crete Algorithm That Ensures Convergence, Transporta- 2, 2002, pp. 198-215.
tion Research Record, 1964, pp. 193-200. http://dx.doi.org/10.1287/trsc.34.2.198.12304
[2] E. W. Dijkstra, A Note on Two Problems on Connexion [11] Z. S. Yang, Basis Traffic Information Fusion Technol-
with Graphs, Numerische Mathematik, Vol. 1, No. 1, ogy and Its Application, China Railway Publish House,
1959, pp. 269-271. Beijing, 2005.
http://dx.doi.org/10.1007/BF01386390
[12] C. Liu, X. L. Meng and Y. M. Fan, Determination of Rou-
[3] M. Sniedovich, Dijkstras Algorithm Revisited: The ting Velocity with GPS Floating Car Data and WebGIS-
Dynamic Programming Connexion, Journal of Control Based Instantaneous Traffic Information Dissemination,
and Cybernetics, Vol. 35, No. 3, 2006, pp. 599-620. The Journal of Navigation, Vol. 61, No. 2, 2008, pp. 337-
[4] R. W. Floyd, Algorithm 97: Shortest Path, Communi- 353. http://dx.doi.org/10.1017/S0373463307004547
cations of the ACM, Vol. 5, No. 6, 1962, p. 345. [13] X. Zhang, Shortest Paths Probability Distribution of
http://dx.doi.org/10.1145/367766.368168 Stochastic Road Network, Master Thesis, 2010.
[5] R. Bellman, On a Routing Problem, Quarterly of Ap- [14] P. B. Mirchandani, Shortest Distance and Reliability of
plied Mathematics, Vol. 16, No. 1, 1958, pp. 87-90. Probabilistic Networks, Computer and Operations Re-
[6] D. Delling, P. Sanders, D. Schultes and D. Wagner, En- search, Vol. 3, No. 4, 1976, pp. 347-355.
gineering Route Planning Algorithms, Algorithmics of http://dx.doi.org/10.1016/0305-0548(76)90017-4
Large and Complex Networks, Vol. 5515, 2009, pp. 117- [15] S. X. Zhang, X. J. Liu and Y. Yang, Dynamic Stochastic
139. http://dx.doi.org/10.1007/978-3-642-02094-0_7 Shortest Path Algorithm, Acta Physica Sinica, Vol. 61,
[7] P. E. Hart, N. J. Nilsson and B. Raphael, A Formal Basis No. 12, 2012, Article ID: 160201.

Copyright 2013 SciRes. JTTs

 Z. Q. SUN ET AL. 265

[16] G. R. Jagadeesh, T. Srikanthan and K. H. Quek, Heuris- [17] C. W. Ahn and R. S. Ramakrishna, A Nondominated
tic Techniques for Accelerating Hierarchical Routing on Sorting Genetic Algorithm for Shortest Path Routing Pro-
Road Networks, IEEE Transactions on Intelligent Trans- blem, IEEE Transactions on Evolutionary Computa-
portation Systems, Vol. 3, No. 4, 2002, pp. 301-309. tion, Vol. 6, No. 6, 2002, pp. 566-579.
http://dx.doi.org/10.1109/TITS.2002.806806 http://dx.doi.org/10.1109/TEVC.2002.804323

Copyright 2013 SciRes. JTTs