Traffic Prediction Based on Improved Neural Network - Semantic Scholar
Journal of Convergence Information Technology Volume 5, Number 9. November 2010
Traffic Prediction Based on Improved Neural Network Cui Jian-ming School of Information Engineering, Chang’an University, Xi’an 710064, China. Shaanxi Road Traffic Detection and Equipment Engineering Research Center, Xi’an 710064,China. Email:[email protected] doi:10.4156/jcit.vol5. issue9.8
Abstract Artificial neural networks and genetic algorithms derived from the corresponding simulation of biology, anatomy. The paper analyzes the advantages and the disadvantages of the artificial neural networks and genetic algorithms. The artificial neural networks and genetic algorithms to be combine in the prediction model. This method is used to predict traffic volume in a road, the accuracy of forecasting results improved significantly. Therefore, this simulation method in traffic prediction have a good prospect.
Keywords: Improved Neural Networks, Traffic, Prediction Models 1. Introduction Traffic prediction is an important item in the transport planning[1]. It is also an indispensable component technology in traffic guidance system. The accuracy of its forecasting results will directly affect the reasonableness of the transport planning, and the correctness of the traffic guidance. Choice of prediction method directly related to the objectives of the forecast results and the accuracy of forecast. Common forecasting methods are often heavy workload, but accuracy is not high. In recent years, artificial neural networks and genetic algorithms are applied gradually. Neural network model is the network level structure in some rules by a multi-neuron connections. It has many advantages: a strong adaptive capacity and ability to learn, nonlinear mapping capabilities, robustness and fault tolerance. But its essence is gradient search method, this easily result in a conspicuous shortcomings of local minimum. Genetic algorithm is adaptive random Global Optimization Algorithm. This algorithm will be every possible problems result from "chromosome" to describe. Through the chromosome replication, exchange and variations to find the optimal calculation. Its operation is the target of a feasible solution for a number of groups, with the nature of the characteristics of parallel computing, so its search speed, but it also shortcomings: First, some issues do not know in advance the value scope of variables; Second, rely entirely on random probability for optimal operation, it is difficult to obtain optimal solutions, a greater impact on human factors. In view of the above their existing problems, In this paper, genetic algorithms and neural network algorithm organically integrate, and make full use of their respective advantages of genetic neural network model constructed using genetic algorithm to train the neural network weights, Overcome the neural network of local Poles problems. Finally, this model is applied to traffic volume forecast for the test results. The results showed that this method is feasible traffic forecasting methods.
2. Basic principles of Genetic Algorithm For the average parameter optimization, can be expressed as: Objective function:
min F f ( x, y, z) ， F R
G( x, y, z) 0
( x, y, z)
Constraints: ， Genetic Algorithm[1] basic idea is: First, the optimization problem for a group of basic feasible solution
( xi , yi , zi )
for a group of binary encoding of the string, (Each string containing multiple
85
Traffic Prediction Based on Improved Neural Network Cui Jian-ming
sub-string, each sub-string in a one or a combination of several known as a gene, also known as chromosome), and then again some of these chromosomes parameter optimization and operations . Then, in order to achieve parameter optimization to operating these chromosomes. Basic processes in genetic algorithm are including replication, exchange and variation, that implementation item process of optimization by imitation use the three methods.
3. Basic principles of neural network The neurons is base unit in Neural networks[2,3]，an n-inputs x i ， wi is connection weight ，u T
is outputs. f (W X ) were called transfer function(activation function). effect is as follows: output change from possible infinite field to finite field, and power processing of nonlinear by imitation the neurons.In this method system, it imitate one processor with multiple inputs and single output. among them: f (W T X ) u f (
n
w x ) i
i
i 1
(1)
T W and X are structure column vector by wi and x i , W Transpose by W Matrix.
4. Based on genetic algorithm neural network weights learning Here, through the repeated use of genetic algorithm optimization weights of neural network, until there is no longer any increase in average. Decoding the parameters of this combination has been fully close to the best combination of parameters. On this basis, re-use neural network algorithm for them to fine-tune(trainings),this method has strong applicability. According to need of the traffic forecasts, using multiple-input and multiple-output feed-forward neural network, the type of neural network algorithm used for BP. Its core ideal is the first given t group initial weights of network, using BP network algorithms training t group weight, and from the minimum and maximum levels corresponding by this t group weights determine the interval value of each weight. Following this, the use of specific coding to generating gene group and using genetic algorithm optimization. Genetic Algorithm chromosome is weights of the neural network. See Figure.1. Evolution Groups
genetic algorithm
Fitness calculation
evaluation system test
training new network
Network after training
training set trainingSet set training
test set testSet set test
Figure 1. Neural Network in the learning process under the Genetic Algorithm
4.1. Coding
86
Journal of Convergence Information Technology Volume 5, Number 9. November 2010
Using real coding, various weight of the neural network to connect a long string according to a certain order, each location corresponding to a network weight in the string. This code means avoiding a coding error in the process of calculation, while reducing the time of the encoding and decoding, and improve the computational efficiency.
4.2. Evaluation function All right value in chromosomes to be assigned to the network structure, training samples for input f
and
output
1 e 2
of
network,
1 e
is
evaluation
function
of
chromosomes.
In
this,
N
∑( y
k
-
y k'
k 1
)2
.
4.3. Initialization process In initial chromosome set, weights of network are determine base on the probability of random, initial weight is random numbers in uniform distribution on from -1.0 to 1.0.
4.4. Genetic Operators The form of genetic operator is multifarious for all application of different, here, using operator of weight of crossover and mutation. Operator of weight crossover, When offspring chromosome select the weights of each , crossover operator selected randomly a number of cross-cutting position form two-parent chromosome and cross computing in this generation of chromosome position. So, the offspring chromosome will contain two parental genes. Operator of weight mutation, input every weights for offspring chromosome, operator select random a value in the distribution of the initial probability, then with the combined weight of the input position.
4.5. Selection Method Here, using method of proportional options. Namely, choice random base on the value of formula (2). popsize
pi =
fi /
f i 1
i
(2)
Which, ( x k , y k ) , (k 1,2,..., N ) are the samples to learning, y k is real output of network. 4.6. Improved Method for Parameters
p
p
Value size of Cross-rate c and mutation m have a great influential for performance of genetic algorithm, under ideal circumstances, with the values of the change of adaptation to change to get
pc
values of
pc
of
pc
and
and
pm
pm
.We used fitness to measure algorithm status of convergence, get lower values
for the solution to high value, the solution was to eliminate. Increasing value
pm
of and when mature before the time of convergence, and speed up the formation of the new entity. with adapt to the changes in value of the fitness, formula of(3)and(4) are formula of
pc
and
pm
.
87
Traffic Prediction Based on Improved Neural Network Cui Jian-ming
f c ) /( f max f ) k ( f p c c max k c f m ) /( f max f ) k ( f p m m max k m
, fc f , fc f
(3)
, fm f , fm f
(4)
kc , k m are constant less than 1;
f c is a large fitness of the two cross individuals . f m is a variation of the individual to individual fitness f max ， f is the greatest fitness and the average fitness in groups. 5. Application of an improved arithmetic on Neural Networks in traffic prediction 5.1. Traffic prediction Using genetic algorithms and neural network forecast the annual average traffic volume of a road section. Passenger volume in 1989 to 2000 input and training as learning samples, look Fig.4 Passenger volume in 2001 to 2004 is the sample set. Neural networks use the three-tier network when the actual forecast, nodes number of input and out is 3,nodes of hide layer is 6,error is 0.001,t is 8,crossover probability pc=0.8,Mutation probability pm=0.01,Max number of iterations is 100, Base on the above parameters and data are as follows Table.1. Table 1. passenger volume annual from 1989 to 2004 [4] year
passenger volume total
year
passenger volume total
1989
791376
1997
1326094
1990
772682
1998
1378717
1991
806048
1999
1394413
1992
860855
2000
1478573
1993
996634
2001
1534122
1994
1092883
2002
1608150
1995
1172596
2003
1587497
1996
1245356
2004
1767453
5.2. Traffic prediction test results According to comparison the results of genetic algorithms, see Table.2, neural network algorithm with improved neural network algorithm. The results can be considered that improved neural network algorithm has a good performance forecast.
88
Journal of Convergence Information Technology Volume 5, Number 9. November 2010
Table 2. Results: genetic algorithms, neural networks algorithm, improved neural network algorithm Algorithms genetic algorithms
neural network algorithm
improved neural network
passenger volume in 2004
passenger volume of prediction by improved neural network in 2008
Values of prediction
1899128
1914293
1694970
1767453
2011153
Average deviation （%）
7.450
8.308
3.101
-
-
items
6. Results and Discussions Improved neural network algorithm optimization calculation model was established, In this paper, the advantages of this model using parallel computing and can be rapid and global search, such a model to resolve that the neural network model has inherent shortcomings, theses are the search slow and the local maturity easy. That model has also used the advantages of the neural network that has a strong ability described problems and has good adaptability for incomplete information, and compensate for the shortcomings of the genetic algorithm coding difficult. And this method has been applied to predict the volume of traffic, the results contrast with forecast of the genetic algorithms and the neural network model, and found that its significantly reduced error in the limited time, have a great degree of increase for forecast accuracy and efficiency. This shows, the predicted method is feasible, neural networks and genetic algorithms organic integration that is an effective tool can be used as predicted the volume of traffic.
7. Acknowledgements The Project was Supported by the Special Fund for Basic Scientific Research of Central Colleges(CHD2010JC035), Chang′an University, by the Special Fund for Basic Research Program of Chang′an University and by the open Fund for Shaanxi Road Traffic Detection and Equipment Engineering Research Center.
8. References [1] CHEN Hong, "An Approach to Traffic Volume Forecasting Based on Ant Colony Neural Network", JDCTA: International Journal of Digital Content Technology and its Applications, Vol. 4, No. 4, pp. 58 ~ 63, 2010X.L. [2] Du, L. Han and L.P. Jiang. “An efficient global optimization Algorithm: Empirical genetic algorithm”, Journal of Beijing Polytechnic University, vol.32,no. 11,pp.992-995,2006 [3] Shifei Ding, Weikuan Jia, Chunyang Su, Xiaoliang Liu, Jinrong Chen, "An Improved BP Neural Network Algorithm Based on Factor Analysis", JCIT: Journal of Convergence Information Technology, Vol. 5, No. 4, pp. 103 ~ 108, 2010 [4] National bureau of Statistics of China.《China Statistical Yearbook 2005》.Beijing, China Statistics Press
89
Traffic Prediction Based on Improved Neural Network Cui Jian-ming School of Information Engineering, Chang’an University, Xi’an 710064, China. Shaanxi Road Traffic Detection and Equipment Engineering Research Center, Xi’an 710064,China. Email:[email protected] doi:10.4156/jcit.vol5. issue9.8
Abstract Artificial neural networks and genetic algorithms derived from the corresponding simulation of biology, anatomy. The paper analyzes the advantages and the disadvantages of the artificial neural networks and genetic algorithms. The artificial neural networks and genetic algorithms to be combine in the prediction model. This method is used to predict traffic volume in a road, the accuracy of forecasting results improved significantly. Therefore, this simulation method in traffic prediction have a good prospect.
Keywords: Improved Neural Networks, Traffic, Prediction Models 1. Introduction Traffic prediction is an important item in the transport planning[1]. It is also an indispensable component technology in traffic guidance system. The accuracy of its forecasting results will directly affect the reasonableness of the transport planning, and the correctness of the traffic guidance. Choice of prediction method directly related to the objectives of the forecast results and the accuracy of forecast. Common forecasting methods are often heavy workload, but accuracy is not high. In recent years, artificial neural networks and genetic algorithms are applied gradually. Neural network model is the network level structure in some rules by a multi-neuron connections. It has many advantages: a strong adaptive capacity and ability to learn, nonlinear mapping capabilities, robustness and fault tolerance. But its essence is gradient search method, this easily result in a conspicuous shortcomings of local minimum. Genetic algorithm is adaptive random Global Optimization Algorithm. This algorithm will be every possible problems result from "chromosome" to describe. Through the chromosome replication, exchange and variations to find the optimal calculation. Its operation is the target of a feasible solution for a number of groups, with the nature of the characteristics of parallel computing, so its search speed, but it also shortcomings: First, some issues do not know in advance the value scope of variables; Second, rely entirely on random probability for optimal operation, it is difficult to obtain optimal solutions, a greater impact on human factors. In view of the above their existing problems, In this paper, genetic algorithms and neural network algorithm organically integrate, and make full use of their respective advantages of genetic neural network model constructed using genetic algorithm to train the neural network weights, Overcome the neural network of local Poles problems. Finally, this model is applied to traffic volume forecast for the test results. The results showed that this method is feasible traffic forecasting methods.
2. Basic principles of Genetic Algorithm For the average parameter optimization, can be expressed as: Objective function:
min F f ( x, y, z) ， F R
G( x, y, z) 0
( x, y, z)
Constraints: ， Genetic Algorithm[1] basic idea is: First, the optimization problem for a group of basic feasible solution
( xi , yi , zi )
for a group of binary encoding of the string, (Each string containing multiple
85
Traffic Prediction Based on Improved Neural Network Cui Jian-ming
sub-string, each sub-string in a one or a combination of several known as a gene, also known as chromosome), and then again some of these chromosomes parameter optimization and operations . Then, in order to achieve parameter optimization to operating these chromosomes. Basic processes in genetic algorithm are including replication, exchange and variation, that implementation item process of optimization by imitation use the three methods.
3. Basic principles of neural network The neurons is base unit in Neural networks[2,3]，an n-inputs x i ， wi is connection weight ，u T
is outputs. f (W X ) were called transfer function(activation function). effect is as follows: output change from possible infinite field to finite field, and power processing of nonlinear by imitation the neurons.In this method system, it imitate one processor with multiple inputs and single output. among them: f (W T X ) u f (
n
w x ) i
i
i 1
(1)
T W and X are structure column vector by wi and x i , W Transpose by W Matrix.
4. Based on genetic algorithm neural network weights learning Here, through the repeated use of genetic algorithm optimization weights of neural network, until there is no longer any increase in average. Decoding the parameters of this combination has been fully close to the best combination of parameters. On this basis, re-use neural network algorithm for them to fine-tune(trainings),this method has strong applicability. According to need of the traffic forecasts, using multiple-input and multiple-output feed-forward neural network, the type of neural network algorithm used for BP. Its core ideal is the first given t group initial weights of network, using BP network algorithms training t group weight, and from the minimum and maximum levels corresponding by this t group weights determine the interval value of each weight. Following this, the use of specific coding to generating gene group and using genetic algorithm optimization. Genetic Algorithm chromosome is weights of the neural network. See Figure.1. Evolution Groups
genetic algorithm
Fitness calculation
evaluation system test
training new network
Network after training
training set trainingSet set training
test set testSet set test
Figure 1. Neural Network in the learning process under the Genetic Algorithm
4.1. Coding
86
Journal of Convergence Information Technology Volume 5, Number 9. November 2010
Using real coding, various weight of the neural network to connect a long string according to a certain order, each location corresponding to a network weight in the string. This code means avoiding a coding error in the process of calculation, while reducing the time of the encoding and decoding, and improve the computational efficiency.
4.2. Evaluation function All right value in chromosomes to be assigned to the network structure, training samples for input f
and
output
1 e 2
of
network,
1 e
is
evaluation
function
of
chromosomes.
In
this,
N
∑( y
k
-
y k'
k 1
)2
.
4.3. Initialization process In initial chromosome set, weights of network are determine base on the probability of random, initial weight is random numbers in uniform distribution on from -1.0 to 1.0.
4.4. Genetic Operators The form of genetic operator is multifarious for all application of different, here, using operator of weight of crossover and mutation. Operator of weight crossover, When offspring chromosome select the weights of each , crossover operator selected randomly a number of cross-cutting position form two-parent chromosome and cross computing in this generation of chromosome position. So, the offspring chromosome will contain two parental genes. Operator of weight mutation, input every weights for offspring chromosome, operator select random a value in the distribution of the initial probability, then with the combined weight of the input position.
4.5. Selection Method Here, using method of proportional options. Namely, choice random base on the value of formula (2). popsize
pi =
fi /
f i 1
i
(2)
Which, ( x k , y k ) , (k 1,2,..., N ) are the samples to learning, y k is real output of network. 4.6. Improved Method for Parameters
p
p
Value size of Cross-rate c and mutation m have a great influential for performance of genetic algorithm, under ideal circumstances, with the values of the change of adaptation to change to get
pc
values of
pc
of
pc
and
and
pm
pm
.We used fitness to measure algorithm status of convergence, get lower values
for the solution to high value, the solution was to eliminate. Increasing value
pm
of and when mature before the time of convergence, and speed up the formation of the new entity. with adapt to the changes in value of the fitness, formula of(3)and(4) are formula of
pc
and
pm
.
87
Traffic Prediction Based on Improved Neural Network Cui Jian-ming
f c ) /( f max f ) k ( f p c c max k c f m ) /( f max f ) k ( f p m m max k m
, fc f , fc f
(3)
, fm f , fm f
(4)
kc , k m are constant less than 1;
f c is a large fitness of the two cross individuals . f m is a variation of the individual to individual fitness f max ， f is the greatest fitness and the average fitness in groups. 5. Application of an improved arithmetic on Neural Networks in traffic prediction 5.1. Traffic prediction Using genetic algorithms and neural network forecast the annual average traffic volume of a road section. Passenger volume in 1989 to 2000 input and training as learning samples, look Fig.4 Passenger volume in 2001 to 2004 is the sample set. Neural networks use the three-tier network when the actual forecast, nodes number of input and out is 3,nodes of hide layer is 6,error is 0.001,t is 8,crossover probability pc=0.8,Mutation probability pm=0.01,Max number of iterations is 100, Base on the above parameters and data are as follows Table.1. Table 1. passenger volume annual from 1989 to 2004 [4] year
passenger volume total
year
passenger volume total
1989
791376
1997
1326094
1990
772682
1998
1378717
1991
806048
1999
1394413
1992
860855
2000
1478573
1993
996634
2001
1534122
1994
1092883
2002
1608150
1995
1172596
2003
1587497
1996
1245356
2004
1767453
5.2. Traffic prediction test results According to comparison the results of genetic algorithms, see Table.2, neural network algorithm with improved neural network algorithm. The results can be considered that improved neural network algorithm has a good performance forecast.
88
Journal of Convergence Information Technology Volume 5, Number 9. November 2010
Table 2. Results: genetic algorithms, neural networks algorithm, improved neural network algorithm Algorithms genetic algorithms
neural network algorithm
improved neural network
passenger volume in 2004
passenger volume of prediction by improved neural network in 2008
Values of prediction
1899128
1914293
1694970
1767453
2011153
Average deviation （%）
7.450
8.308
3.101
-
-
items
6. Results and Discussions Improved neural network algorithm optimization calculation model was established, In this paper, the advantages of this model using parallel computing and can be rapid and global search, such a model to resolve that the neural network model has inherent shortcomings, theses are the search slow and the local maturity easy. That model has also used the advantages of the neural network that has a strong ability described problems and has good adaptability for incomplete information, and compensate for the shortcomings of the genetic algorithm coding difficult. And this method has been applied to predict the volume of traffic, the results contrast with forecast of the genetic algorithms and the neural network model, and found that its significantly reduced error in the limited time, have a great degree of increase for forecast accuracy and efficiency. This shows, the predicted method is feasible, neural networks and genetic algorithms organic integration that is an effective tool can be used as predicted the volume of traffic.
7. Acknowledgements The Project was Supported by the Special Fund for Basic Scientific Research of Central Colleges(CHD2010JC035), Chang′an University, by the Special Fund for Basic Research Program of Chang′an University and by the open Fund for Shaanxi Road Traffic Detection and Equipment Engineering Research Center.
8. References [1] CHEN Hong, "An Approach to Traffic Volume Forecasting Based on Ant Colony Neural Network", JDCTA: International Journal of Digital Content Technology and its Applications, Vol. 4, No. 4, pp. 58 ~ 63, 2010X.L. [2] Du, L. Han and L.P. Jiang. “An efficient global optimization Algorithm: Empirical genetic algorithm”, Journal of Beijing Polytechnic University, vol.32,no. 11,pp.992-995,2006 [3] Shifei Ding, Weikuan Jia, Chunyang Su, Xiaoliang Liu, Jinrong Chen, "An Improved BP Neural Network Algorithm Based on Factor Analysis", JCIT: Journal of Convergence Information Technology, Vol. 5, No. 4, pp. 103 ~ 108, 2010 [4] National bureau of Statistics of China.《China Statistical Yearbook 2005》.Beijing, China Statistics Press
89
